Aerosol forecasting is important for air-quality management, health risk assessment and climate change mitigation1,2. However, it is more complex than weather forecasting, owing to the interactions between aerosol physicochemical processes and atmospheric dynamics, resulting in high uncertainty and computational costs3,4. Here we develop a machine-learning-driven Global Aerosol-Meteorology Forecasting System (AI-GAMFS), which provides reliable 5-day, 3-hourly forecasts of aerosol optical components and surface concentrations. AI-GAMFS combines a vision transformer and U-Net in a backbone network, robustly capturing the complex aerosol-meteorology interactions via global attention and spatiotemporal encoding. Trained on 42 years of aerosol reanalysis data and initialized with Global Earth Observing System Forward Processing (GEOS-FP) analyses, AI-GAMFS delivers operational 5-day forecasts in 1 minute. Evaluation with independent ground-based observations suggests improved performance compared with the Copernicus Atmosphere Monitoring Service5 and regional dust models6-9 in forecasting aerosol optical depth and dust components. Compared with GEOS-FP10, it has a lower root-mean-square error for global aerosol optical depth, with comparable dust forecasting skill and improved surface aerosol component forecasts over the USA and China. Our results provide a step forward in leveraging machine learning to refine aerosol forecasting and may help warn against aerosol pollution events such as dust storms and wildfires.

Smoke-related PM2.5 is the primary air quality concern during the summer in Alaska, yet accurate forecasting remains a major challenge. In this study, we use machine learning (ML) techniques to improve smoke forecasts from NOAA's HRRR-Smoke model. We find that the model underestimates surface PM2.5 by a factor of up to five during the wildfire season in Alaska. We evaluated Random Forest (RF), one-dimensional, and two-dimensional convolutional neural network (CNN1D and CNN2D) models. Among them, CNN1D performed the best, reducing the underestimation factor to two or less. Analysis of the relationships between key predictors, such as Surface and Vertical Smoke and calibrated PurpleAir observations, suggests that errors in the vertical distribution of smoke are a primary source of underestimating bias. Atmospheric sounding data further show that the HRRR-Smoke model fails to capture daytime temperature inversion layers during wildfire events. This bias is likely caused by missing fire radiative power (FRP) detections under heavy smoke or cloudy conditions, which leads to low smoke concentrations and under-represented radiation feedback necessary to maintain near-surface inversion. Although accurately representing physical and chemical processes in models remains highly challenging, our results demonstrate ML offers an effective approach to improving daily surface PM2.5 forecasts in wildfire-prone regions like Alaska.

Accurate forecasting of fine particulate matter (PM2.5) remains a global challenge due to spatial gaps, data imbalance, and limited representation of extreme events. This study presents an enhanced Deep Imbalanced Regression (DIR) framework that integrates NASA’s GEOS-FP forecasts with global ground-based PM2.5 observations using a Temporal Convolutional Network (TCN) and a Residual Mixture-of-Experts (ResMoE) architecture. The model was trained on 378,000 samples (2021–2025) from U.S. Embassy AirNow sites and OpenAQ sensors, increasing geographic diversity. To address the imbalance, Label Distribution Smoothing (LDS) and weighted loss were applied, while ResMoE adaptively routed samples to specialized experts across meteorological-aerosol regimes. This configuration achieved strong performance (R² = 0.88, MSE = 23.4, MAE = 2.86 µg/m³) and generalized well across polluted and clean regions, including unseen sites. During the May 2025 Minnesota wildfire, the model captured both temporal evolution and peak magnitude missed by the TCN baseline, demonstrating improved responsiveness to extreme events. Uncertainty quantification and sensitivity analysis confirm model consistency. Beyond forecasting, the framework enables spatiotemporally consistent PM2.5 reconstruction for exposure assessment and policy analysis in data-scarce regions. This study provides a scalable and interpretable pathway for next-generation global air-quality forecasting.

Sun-sky radiometer (model POM-01) is commonly used for studying aerosol optical and physical properties at selected aerosol-specific channels. Although the instrument is equipped with a precipitable water vapor (PWV) channel at 940 nm, the inbuilt software does not provide a tool for estimation of PWV. Hence, the current study adopted a new methodology to estimate PWV from three high-altitude (> 3400 m MSL) sites, Hanle, Merak, and Leh, located in Ladakh, India. The retrieval algorithm focuses on the precise estimation of the calibration constant (V0) and coefficients a and b using modified Langley plots in two different methods. The estimated average value of b is 0.59 +/- 0.09 which is very close to those commonly used in global studies. Further, the estimated V0 and b values from both methods are found to be similar, which may be due to the advantages of the dry and high-altitude environment, where the annual total column water vapor is typically less than 6 mm. The estimated PWV using observations at selected full clear and stable atmospheric conditions compares well with satellite, AERONET, GPS, reanalysis and empirical model data with correlation coefficient varying from 0.91 to 0.97. Further, the estimated propagated root mean square error (rmse) varies from 0.37 mm to 2.58 mm. These results indicated that sun-sky radiometer derived PWV showed good consistency with the derived PWV from independent data sources at the three sites.

Abstract. Cloud screening algorithms have always been a critical component of Aerosol Robotic Network (AERONET) aerosol optical depth (AOD) Level 1.5 and 2.0 product. The initial cloud screening algorithm in the Version 1 and 2 database was semi-automatic and required involvement of human analyst to finalize the results. It became fully automatic in Version 3 (V3) due to employing information on the angular shape of sky radiances measured in aureole (curvature algorithm). Although efficient, the curvature algorithm is threshold based and fails to detect clouds when its parameters are beyond the corresponding pre-determined thresholds. This is especially noticeable at high latitudes where the size of ice crystals in cirrus clouds are sometimes relatively small and therefore comparable in size to aerosols. It is shown that additional information can be extracted from analysis of the smoothness of diurnal variability of sky radiances measured at the 3.3-degree scattering angle. This measurement is a part of so-called curvature scan (CCS), which takes measurements from 3 to 7.5 degrees scattering angle with 0.3-degree steps after each measurement of AOD. The analysis of the diurnal variability of CCS (3.3) for cloud-free conditions shows relatively smooth temporal dependencies, which can be fitted by polynomials with high correlation coefficients while in conditions almost completely dominated by clouds, the temporal variability is completely random. For partially cloudy days, the two main features are observed: relatively smooth aerosol signature and irregular spikes due to clouds. The new technique is proposed that employs the smoothness of the diurnal variability of CCS(3.3) scan as a criterion of the cloud free conditions. In the case when both features are present, the idea of the new algorithm is to remove irregular spikes due to clouds while keeping smooth part due to aerosols intact. The new algorithm detects spikes associated with clouds by comparing magnitudes of CCS(3.3) at neighboring time stamps through calculating their first differences (FD). This algorithm was applied to the CCS(3.3) measurements taken at several AERONET sites. The results were analyzed in terms of net change in Angstrom exponent (AE) as well as number of AOD measurements. The analysis showed the algorithm performs satisfactorily at AERONET sites dominated by fine mode aerosols, however at sites dominated by dust, the algorithm removes a big fraction of cloud-free observations. The issue was corrected by introducing an additional cloud screening parameter. It is based on observation of the different rate in changing of AE with iterations for cloud-free and cloudy conditions with much higher rate in the former case. The new parameter was selected as a slope of the linear regression between integration number and the value of AE after the corresponding iteration. Algorithm disregards FD algorithm results if the slope is smaller than certain threshold value. Finalizing the FD algorithm threshold setting as well as evaluation of the algorithm performance is done by using independent cloud detection information available from Micro-Pulse Lidar Network (MPLNET) data. The AERONET and MPLNET data were time and space collocated with additional averaging over one hour period. The comparison showed that, on average, the FD algorithm outperformed V3 L1.5 by about 0.02 in Mathews Correlation Coefficient (MCC), suggesting consistent improvement in overall cloud detection accuracy. Additional analysis performed in terms of MCC metrics also showed that the FD algorithm achieves a more balanced and accurate classification of clouds vs clear.

This special issue addresses emerging technologies and future directions in air quality research by integrating spaceborne observations with in situ measurements, data fusion frameworks, and advanced computational techniques. The collective findings of the contributing studies offer a valuable resource for researchers, practitioners, and policymakers seeking to understand and quantify air pollution across diverse environments. The methodologies presented across these papers establish a foundation for identifying pollution sources and characterizing pollutant transport and transformation processes at the regional scale, supporting the stabilization of air quality management systems. To reduce the health burden of ambient air pollution, the contributing authors collectively underscore the need to raise awareness around reducing anthropogenic emissions and to advance space-driven data fusion systems, including expanding monitoring infrastructure, operationalizing AI/ML-driven analytical pipelines, implementing science-informed emission control policies, and fostering meaningful community engagement.

Delhi is a critical global hotspot for Nitrogen Dioxide (NO2), fueled by dense traffic and industrial hubs. Effective monitoring requires high-resolution data to capture local gradients, yet long-term analysis is often hindered by the coarse resolution of historical Ozone Monitoring Instrument (OMI) products (similar to 25 km). This study develops deep learning models to downscale OMI Multi-Decadal Nitrogen Dioxide and Derived Products from Satellites (MINDS) data specifically for the Delhi metropolitan region. By integrating high-resolution Tropospheric Monitoring Instrument (TROPOMI) data and Central Pollution Control Board (CPCB) ground observations, we bridge the gap between historical records and modern precision. Optimized using the 2024 monitoring cycle, our architecture accurately reconstructs NO2 spatial patterns across extreme seasonal shifts, from post-monsoon pollution peaks to summer transitions. This approach provides a high-fidelity framework for analyzing long-term urban air quality trends at a granular scale. To achieve this, we implemented and evaluated two models: 1) Gated Fusion Downscaling Predictor (GFDP) and 2) a hybrid model that comprises Inverse Distance Weighting with Deep Neural Networks (IDW + DNN). Quantitative evaluation against withheld ground-based data confirmed high model accuracy, yielding a normalized root mean square error (NRMSE) of 0.05, a mean bias of 0.03, and a normalized mean absolute error (NMAE) of 0.04. This advancement is essential for monitoring long-term NO2 trends, provides improved spatial characterization of NO2 distribution, enabling more accurate identification of localized pollution hotspots and supports better interpretation of air quality patterns in complex urban environments.

The Met Office operates a ground based operational network of nine polarisation Raman lidars (aerosol profiling instruments) and sun photometers (column integrated information) across the United Kingdom (UK). An aerosol classification scheme using supervised machine learning has been developed. The concept of Mahalanobis (~normalized) distance to identify the aerosol type from individual Aerosol Robotic Network (AERONET) measurements including Extinction Angstrom Exponent, Absorption Angstrom Exponent, Single Scattering Albedo and Index of refraction is used for a subset of AERONET stations around the globe of known main aerosol types (training set). The aerosol types so far include marine, urban industrial, biomass burning and dust. The relation of particle linear depolarisation ratio (PLDR) and lidar ratio (LR) from the Raman lidar is used in synergy to validate the particle type.

ABSTRACT This product provides MERRA‐2 bias‐corrected global hourly surface total PM2.5 mass concentration with the exact horizontal spatial resolution as MERRA‐2, covering a temporal range from 2000 to 2024. It is derived using a machine learning (ML) approach with a convolutional neural network (CNN) method. It is specifically developed for the NASA Health and Air Quality Applied Sciences Team (HAQAST). The dataset consists of two parameters: MERRA2_CNN_Surface_PM25 and QFLAG. MERRA2_CNN_Surface_PM25, a 3‐dimensional variable (time, latitude, longitude), represents the surface PM2.5 concentrations in μg/m3. QFLAG denotes the quality of data at each grid point, where four indicates the highest quality and 1 indicates the lowest quality. It is recommended to use QFLAG values of 3 and 4 for quantitative analysis.

The use of low-cost sensors (LCS) for air quality monitoring has grown rapidly across a wide range of groups, including community and citizen scientists, academic researchers, environmental agencies, and the private sector. Traditional air monitoring conducted by regulatory agencies relies on expensive, regulatory-grade instruments that require frequent maintenance and rigorous quality control procedures. In contrast, the low purchase price, minimal operating costs, user-friendly design, and open data accessibility have significantly contributed to the widespread adoption of LCS. Over the past decade, hundreds of studies have proposed diverse calibration strategies to tailor LCS performance to specific project needs. This study examines the role of PM2.5 sensors in monitoring air quality across contrasting environments and highlights the importance of inter-sensor consistency. We evaluate PurpleAir (PA) PA-II sensors against regulatory-grade Federal Equivalent Method (FEM) PM2.5 instruments and develop calibration algorithms to improve data accuracy. Calibration deployments were conducted for 2–4 weeks in Raleigh, North Carolina, and Delhi, India, to assess sensor behavior under different aerosol loadings and environmental conditions. The goal of this effort is to create a robust calibration model that uses PA-measured parameters, PM2.5, temperature, and relative humidity as inputs to generate bias-corrected hourly PM2.5 values. The model relies on concurrent FEM PM2.5 measurements as the reference data during calibration development. Multiple statistical and machine-learning approaches were applied to produce a regional calibration model. Our results show that, with proper calibration, PA sensors can provide bias-corrected PM2.5 estimates within 12

A multi-year analysis of aerosol optical depth (AOD, tau) and & Aring;ngstr & ouml;m exponent (alpha) was conducted using ground-based photometer data from 15 Arctic and 11 Antarctic sites. Extending the dataset of through December 2024, the study incorporates stellar and lunar photometric observations to fill data gaps during the polar night. Daily mean values of tau at 0.500 mu m and alpha (0.440-0.870 mu m) were used to derive monthly means and seasonal histograms.In the Arctic, persistent haze events in winter and early spring lead to peak tau values. A decreasing trend in Arctic tau suggests the impact of European emission regulations, while biomass-burning aerosols are becoming more significant. In Antarctica, tau increases from the plateau to the coast. Fine-mode aerosols dominate in summer-autumn, while coarse-mode particles are more prevalent in winter-spring. Shipborne photometer data align well with ground-based measurements, confirming the reliability of mobile observations.Trend analyses using the Mann-Kendall test and Theil-Sen regression indicate a significant negative trend in tau at Andenes (-2.43 % per year), likely driven by reduced anthropogenic emissions. Antarctic stations such as Syowa and South Pole show positive trends (+3.84 % and +3.54 % per year), though these are subject to uncertainties from data limitations and instrument changes.This work contributes to the Polar-AOD network (https://polaraod.net/, last access: 15 May 2025), enhancing the understanding of aerosol variability and long-term trends in polar regions while promoting open data access for the scientific community.

Abstract. The NASA airborne Arctic Radiation-Cloud-aerosol-Surface-Interaction Experiment (ARCSIX) collected a unique data set providing a near-simultaneous characterization of radiative fluxes, surface, cloud, and aerosol particle properties to address science questions on the surface radiation budget, the processes governing the cloud lifecycle, atmospheric composition, and the interactions between the surface and atmosphere. The overarching goal of ARCSIX was to quantify the contributions of surface, clouds, aerosol particles, and precipitation to summer sea ice melt. ARCSIX consisted of two deployments in 2024 (Spring: 2024-05-28 through 2024-06-13 and Summer: 2024-07-25 through 2024-08-15) to capture pre- and post-melt conditions. ARCSIX provided coordinated remote sensing and in situ sampling using three aircraft in a high-flyer/low-flyer configuration. The NASA G-III served as the high-flying remote sensing platform with two lower flying in situ and near-target remote sensor observing platforms, NASA P-3B and SPEC Inc. Learjet. ARCSIX data are well-suited to improve satellite remote sensing capabilities in the Arctic. ARCSIX included an array of sea ice mass balance buoys deployed in the Lincoln Sea that were regularly overflown during the campaign. ARCSIX research flights spanned the Baffin Bay, Lincoln Sea, west and north of the Canadian Archipelago, and the Greenland north and northeast coasts. During the spring deployment, 19 research flights took place covering 114 flight hours: 10 flights and 68 hours by the P-3B and nine flights and 46 hours by the G-III. During summer, 24 research flights covered 136 flight hours: nine flights and 75 hours by the P-3B, five flights and 26 hours by the G-III, and 10 flights and 35 hours by the Learjet. A total of 13 coordinated flights with 2+ aircraft were carried out. This paper describes the ARCSIX flight strategy, instrumentation, and data set access, and usage details. ARCSIX data are publicly available at https://doi.org/10.5067/SUBORBITAL/ARCSIX/DATA001.

This study estimates ground-level fine particulate matter (PM2.5) concentrations using geostationary satellites-derived Aerosol Optical Depth (AOD) and radiance measurements and meteorological parameters from the High-Resolution Rapid Refresh (HRRR) model, with AirNow PM2.5 measurements over the contiguous United States (CONUS). A Deep Neural Network (DNN) was adopted and compared with other machine learning (ML) models (i.e., Random Forest and Light Gradient-Boosting Machine) to estimate surface PM2.5 concentrations. The DNN model (without the tropospheric emissions: monitoring of pollution (TEMPO); 1 year) estimated PM2.5 with an interquartile range (IQR) of 4.32 μg/m3, and outperformed ML models, with up to 44.68% better index of agreement (IOA) and 45.28% smaller relative root-mean-square error (rRMSE), particularly in high PM2.5 cases. The hourly estimated PM2.5 closely matched the observed PM2.5 in both temporal trend and spatial distribution across the eastern CONUS. ML modeling was further enhanced to include TEMPO Level 1b (L1b) data. The DNN model with TEMPO improved performance, with an 8% higher R 2 and a 25% lower rRMSE than the DNN model without TEMPO. The more significant improvement was seen during high smoke events using the TEMPO data. For the first time, we demonstrate the use of TEMPO L1b spectrally resolved radiances data to capture high PM2.5 concentrations during the wildfire events, enhancing our understanding of PM2.5 dynamics. This study provides a framework to integrate data from multiple geostationary satellites with HRRR model outputs to estimate surface air quality at high temporal resolution.

This study aims to provide the first analysis of aerosol optical properties, radiative forcing, and source identification over Birkat al Mouz, Oman, using Aerosol Robotic Network (AERONET) data from December 2022 to November 2024. We analyzed Aerosol Optical Depth (AOD), Angstrom Exponent (AE), Single Scattering Albedo (SSA), aerosol radiative forcing (RF), and performed Concentration-Weighted Trajectory (CWT) analysis to identify aerosol transport pathways and sources. The highest aerosol loading (AOD = 0.49 ± 0.15) occurred in summer, with the lowest (0.17 ± 0.08) in winter. AE values (maximum 0.94 ± 0.20 in winter, minimum 0.42 ± 0.17 in summer) indicated coarse-mode aerosol dominance. Seasonal SSA values were highest in summer (0.95), confirming significant dust aerosol influence. Surface RF averaged − 43.81 W m−2, atmospheric RF was 27.04 W m−2, and aerosol-induced heating reached 0.74 K day−1. CWT analysis revealed the Horn of Africa, and arid regions of the Arabian Peninsula as major aerosol sources. Seasonal aerosol variations in Birkat al Mouz are predominantly driven by dust aerosols transported from remote regions, highlighting their significant role in regional climate forcing.

Abstract Estimating surface‐level fine particulate matter from satellite remote sensing can expand the spatial coverage of ground‐based monitors. This approach is particularly effective in assessing rapidly changing air pollution events such as wildland fires that often start far away from centralized ground monitors. We developed Deep Neural Network (DNN) algorithm to improve hourly PM2.5 estimates informed by GOES‐R; meteorology forecasts, and PM2.5 observations from AirNow. The surface‐satellite‐model collocated data sets for the period of 2020–2021 were used to assess the biases in GOES‐GWR PM2.5 (only operationally available data set) against AirNow measurements at hourly and daily scales. Then a DNN based bias correction algorithm is used to improve the accuracies of GOES‐GWR PM2.5. The DNN uses GOES‐GWR PM2.5, GOES‐R aerosol parameters, and HRRR meteorological fields as input and AirNow PM2.5 is used as target variable. The application of DNN reduced the PM2.5 biases as compared to GOES‐GWR estimates. RMSE was also reduced to 6.55 μg/m3 from 8.72 μg/m3 in GOES‐GWR estimates. The DNN model was also evaluated on two sets of independent data sets for its robustness. In the first independent data set for the first half of 2020, ∼89% of stations show an increase in correlation (r) and, ∼76% and ∼62% of stations show a reduction in bias. The IOA and r for the independent data were 0.77 and 0.61 (GWR: 0.68 and 0.53) and RMSE was 4.48 μg/m3 (GWR = 6.13 μg/m3) for the same period. The algorithm will be operationally deployed by NOAA and US‐EPA to estimate surface level PM2.5 from satellite derived Aerosol optical depth.

Accurate forecasting of PM2.5 (particulate matter <= 2.5 mu m) is essential for effective air quality management, particularly in urban areas such as Delhi, which frequently experience severe pollution episodes. This study evaluates the predictive capabilities of regional and global forecasting models for PM(2.5 )concentrations and the associated Air Quality Index (AQI) in Delhi, India. A multi-model assessment was conducted using three regional models (WRF-Chem, SILAM, and DM-Chem) and four global models (IFS, GEOS-FP, GEFS-Aerosols, and the machine learning-based GEOS-ML). Forecasts from these models were validated against hourly in situ measurements from 39 Central Pollution Control Board (CPCB) stations in Delhi. Results revealed that the Air Quality Early Warning System (AQEWS) based on WRF-Chem exhibited the highest predictive accuracy (Performance Index, PI = 87), with minimal deviations from observations. The GEOS-ML model (PI = 70) effectively captured key variations using a machine learning approach. DM-Chem (330 m: PI = 69, 1.5 km: PI = 61) showed reasonable agreement, whereas IFS (PI = 60), GEOS-FP (PI = 52), and GEFS-Aerosols (PI = 47) captured broader trends with varying accuracy. SILAM (PI = 58) exhibited notable discrepancies during high-pollution events. This study underscores the need for rigorous evaluation of forecasting systems to enhance air quality prediction in polluted urban environments such as Delhi. Identifying the most reliable models supports data-driven decision-making for air pollution mitigation and public health protection. Plain Language Summary Air pollution in Delhi, India, significantly worsens during the post- monsoon and winter seasons due to local human activities and agricultural residue burning in surrounding areas. This study evaluates the effectiveness of various air quality forecasting models in predicting PM2.5 levels and the Air Quality Index (AQI) in Delhi. It compares regional models, which provide detailed local forecasts, and global models, which cover larger areas but often miss localized pollution spikes such as those during Diwali. The findings indicate that the WRF-Chem model, a regional model, is the most accurate, particularly for short- term pollution events. In contrast, global models typically underperform in these scenarios. An ensemble approach, combining multiple models, improves the accuracy of predictions. This study highlights the importance of using a diversified modeling approach to accurately predict and manage air quality in urban environments such as Delhi, providing crucial insights for public health protection and policy-making.

The present study performed classification global aerosols based on particle linear depolarization ratio (PLDR) and single scattering albedo (SSA) provided from AErosol RObotic NETwork (AERONET) Version 3.0 and Level 2.0 inversion products of 171 AERONET sites located in six continents. Current methodology could distinguish effectively between dust and non-dust aerosols using PLDR and SSA. These selected sites include dominant aerosol types such as, pure dust (PD), dust dominated mixture (DDM), pollution dominated mixture (PDM), very weakly absorbing (VWA), strongly absorbing (SA), moderately absorbing(MA), and weakly absorbing (WA). Biomass-burning aerosols which are associated with black carbon are assigned as combinations of WA, MA and SA. The key important findings show the sites in the Northern African region are predominantly influenced by PD, while south Asian sites are characterized by DDM as well as mixture of dust and pollution aerosols. Urban and industrialized regions located in Europe and North American sites are characterized by VWA, WA, and MA aerosols. Tropical regions, including South America, South-east-Asia and southern African sites which prone to forest and biomass-burning, are dominated by SA aerosols. The study further examined the impacts by radiative forcing for different aerosol types. Among the aerosol types, SA and VWA contribute with the highest (30.14 +/- 8.04 Wm-2) and lowest (7.83 +/- 4.12 Wm-2) atmospheric forcing, respectively. Consequently, atmospheric heating rates are found to be highest by SA (0.85 K day-1) and lowest by VWA aerosols (0.22 Kday-1). The current study provides a comprehensive report on aerosol optical, micro-physical and radiative properties for different aerosol types across six continents.