Drone measurements reveal high near-surface urban haze

Abstract

Post-winter haze events in Delhi, India, comprise great air quality challenges, yet remain poorly understood due to limited measurements of vertical profiles of particulate matter (PM) concentrations. This study employs a drone-mounted PM low-cost sensor (PM-LCS) with an optimized sampling system to capture vertical PM_2.5 profiles during March 2021. Elevated PM_2.5 concentrations (160 µg/m³) were observed at an altitude of 100 m, being 60% higher than ground level. Vertical profiles of the PM₁/PM_2.5 ratio under humid conditions (RH > 70%), showed that haze formation is likely driven by hygroscopic inorganic aerosols. Comparison with model simulations showed significant underestimation of PM_2.5 (−52.6 ± 5.5%) during morning haze episodes, coinciding with a dry bias in modeled RH (−30.1 ± 8.3%). During non-hazy episodes, PM_2.5 underestimation decreased to 10.8 ± 1.2% with a minimal RH bias. This suggests that the dry bias of the model limits its ability to simulate aerosol hygroscopic growth. Overall, our findings demonstrate that drone-mounted PM-LCS provides a valuable vertical air quality assessment tool.

Introduction

Air pollution remains one of the most pressing environmental challenges, significantly impacting public health and climate^1,2. At the same time, severe haze episodes have been observed in highly polluted cities in various parts of the world, including Delhi^3,4,5,6,7, Beijing^8,9,10, Cairo ^11,12,13 and Paris¹⁴. Haze pollution, resulting from high ambient particulate matter (PM) (especially fine particles having sizes less than 2.5 µm; PM_2.5) concentration, reduces visibility and poses a serious threat to climate, air quality, and health^1,15,16. The evolution of haze generally exhibits distinctive secondary aerosol formation, which is linked to several chemical and physical processes, including new particle formation and aerosol growth driven by photochemistry and nighttime chemistry during the clean stage and aqueous chemistry during the transition stage¹⁷. In addition, the water uptake of aerosol particles is an important driver for haze formation³. Additionally, the aerosol concentration and haze formation strongly depend on conducive weather conditions such as the boundary layer height^18,19,20,21.

During post-monsoon and winter seasons, Delhi’s National Capital Region (NCR) experiences severe hazy events that impact climate, health, and economy^4,22. As a result, a number of recent ground-based measurements have been conducted in order to understand the composition, sources, and formation mechanisms of severe haze episodes in northern India^{3,4,5,6,23,24}. Despite this, information on post-winter haze events in Delhi is still limited^25,26. Additionally, the vertical pollutant information during the post-winter haze remains extremely negligible. Most of the studies performed in India are either based on satellite-derived aerosol optical depth (AOD) data, or focused on total ground-based PM_2.5 concentration^27,28,29. The vertical distribution of air pollutants and meteorological variables governs their transport, transformation, and surface-level impacts, yet remains a critical uncertainty in atmospheric sciences³⁰. Vertical profiling resolves this gap by quantifying the three-dimensional distribution of pollutants and meteorological drivers, which is a prerequisite for constraining their surface-level impacts. Additionally, the vertical information enables improved scaling of satellite-derived AOD to ground-level PM_2.5, or LIDAR measurements³¹ a pivotal conversion for top-down air quality analysis but inherently limited by assumptions in aerosol layer height, mixing dynamics, and distribution^32,33. Vertical measurements over the region of investigation in the lower atmospheric boundary layer (ABL) are extremely limited^{34,35,36,37,38}, and mainly reported during the end-of winter season. Understanding the vertical distribution of PM in the lower boundary layer (both in winter and post-winter months), where most pollution processing occurs, is critical for refining air quality models such as Weather Research Forecasting coupled with chemistry (WRF-Chem) and improving predictive capabilities³⁹. Traditional ground-based air quality monitoring networks provide valuable but spatially limited data, often missing crucial information about vertical variations in pollutant concentrations. In contrast, emerging drone-based measurement platforms equipped with particulate matter low-cost sensors (PM-LCS) offer a novel and flexible approach to capturing the vertical structure of air pollution during haze episodes and under normal atmospheric conditions^31,40,41,42.

LCSs have gained significant attention due to their affordability, portability, and high spatial-temporal resolution⁴³. Unlike regulatory-grade monitors, which are expensive and require complex infrastructure, LCSs provide an opportunity for widespread deployment, allowing for enhanced data collection in urban and rural areas alike^44,45. These sensors work on various principles such as optical scattering, electrochemical detection, and metal oxide sensing, enabling the measurement of multiple air pollutants, including PM₁, PM_2.5, NO_x, and CO etc. However, despite their advantages, LCSs face challenges in terms of calibration, sensor drift, and interference from environmental conditions, necessitating rigorous validation and correction procedures^46,47,48,49.

Several studies have demonstrated the efficacy of LCSs in air pollution research using drones, particularly for monitoring ambient particulate matter and gases^{40,41,42,50,51,52,53,54}. Despite their advantages, using LCS on-board drones presents challenges, including interference from rotor-induced turbulence and the influence of meteorological factors such as humidity⁵⁵. Drone-based inter-comparison evaluations have been conducted to mitigate these issues, focusing on optimizing sensor placement, flight patterns, and data processing techniques^42,50,54. In addition, dedicated facilities where LCSs are tested under low-pressure/temperature conditions, typically found at higher altitudes, and integrated on-board drones have also been developed^56,57. The turbulence created by drone rotors can affect PM concentration readings, leading researchers to develop sampling strategies that minimize these disturbances, such as designing a sampling inlet that can be placed vertically or horizontally on-board the drone, positioning of sensors at a certain height away from the propeller wash, and using shielding techniques^42,50,53,58. A recent study has highlighted these challenges in detail and provided necessary measures for minimizing these issues⁴⁷.

Drone-based LCS measurements have also shown promise in validating numerical models such as WRF-Chem³⁹. The lower boundary layer, where pollutant dispersion and transformation predominantly occur, is a critical zone for model evaluation⁵⁹. Accurate observations of PM mass concentrations at different altitudes provide essential constraints for improving model parameterizations, ensuring better predictions of pollution transport and haze formation. Integrating these measurements into modeling frameworks allows researchers to refine simulations of boundary layer processes, ultimately leading to enhanced forecasting of air quality events and better-informed mitigation strategies.

In this study, we deployed a drone equipped with a low-cost PM sensor to investigate vertical variations in air pollutant concentrations during different times of the day (morning, noon, and evening), particularly focusing on hazy and non-hazy mornings in Delhi. We present vertical profiles of the PM₁/PM_2.5 ratio as a proxy to infer the potential haze formation sources. We also compare PM_2.5 mass concentration measurements with WRF-Chem simulated PM_2.5 values during the sampling period. Our analysis demonstrates the effectiveness of LCSs for air quality assessments and model validations while overcoming key methodological challenges, including minimizing the impacts of rotor and humidity, and inter-sensor comparisons^39,42,53. Insights gained from these measurements pave the way for integrating drone-based observations into routine air quality monitoring and atmospheric modeling frameworks, enhancing our ability to manage and mitigate pollution episodes.

Results

Probable drivers of near-surface early morning haze formation at Delhi

During 11-23 March 2021, the near-surface composition of the atmospheric aerosol over Delhi exhibited distinct diurnal variations (see Fig. 1), with mass concentrations of organics above 40 µg/m³ during the early morning hours (05:00-08:00 LT) later dropping to less than 20 µg/m³ by noon. Organic aerosols constituted the dominant fraction (40-50%), yet the role of inorganics, particularly chloride (episodic chloride peaks) (20-30%), and the rest of the inorganics such as nitrate (10-15%), sulfate (5-10%), and ammonium (5-10%), found to be critical in driving the observed haze episodes. The increase in the mass concentrations of chloride and ammonium during the night, followed by their rapid depletion post-sunrise, highlights the influence of nocturnal aerosol chemistry and condensational processes, where high RH (above 80%) and enhanced ALWC (40-50 µg/m³) could have promoted multiphase chemistry and secondary aerosol growth^3,6. Chloride, typically present at lower concentrations (compared to other aerosol chemical constituents), exhibited distinct early-morning spikes exceeding ~20 µg/m³, temporally coincident with peak haze formation. These sudden chloride spikes, observed between 04:00 and 07:00 LT, suggest an active role in rapid aerosol-phase transitions, likely through heterogeneous reactions, and increasing aerosol hygroscopicity^3,6. The simultaneous increase in equivalent black carbon (eBC, ~10-20 µg/m³) (coinciding elevated chloride values) pointed to primary emissions from combustion sources, including biomass and solid fuel burning, reinforcing its role in providing a surface for secondary aerosol formation⁴. Elevated NO₂ levels (~100 µg/m³) during morning rush hours (06:00-09:00 LT) coincided with a suppressed planetary boundary layer (PBL height <200 m), restricting vertical mixing and facilitating pollutant entrapment near the surface. Simultaneous drone-based profiling conducted from 18 to 23 March 2021 revealed higher PM concentrations at high altitudes, especially during hazy mornings. In one case, PM_2.5 measurements by a drone-mounted PM-LCS reached 160 µg/m³ around 100 m on 18 March 2021, highlighting the importance of a shallow nocturnal boundary layer in near-surface haze persistence. The interplay among secondary aerosol formation, hygroscopic growth, suppressed dispersion, and the inorganics, particularly chloride and nitrate, compounded by weak wind conditions (<2 m/s), could altogether modulate early-morning haze intensity⁶. The enhanced ALWC during these episodes further suggests a strong feedback mechanism, where multiphase chemistry likely accelerates secondary aerosol production, exacerbating haze conditions before boundary layer expansion and facilitating dispersion later in the day⁶. These findings emphasize the need for a more detailed understanding of aerosol chemistry in haze formation, especially at higher altitudes, particularly its episodic enhancement and interactions with other secondary species, which remain underexplored in the context of South Asian megacities.

Fig. 1: Surface and reanalysis observations during 11 to 23 March 2021 at the Indian Institute of Technology, Delhi.

During the first days of the campaign, ground-based and in-flight comparisons were performed. Later in the campaign, vertical profiling of air pollutants using the drone platform is highlighted by the black dotted line box (18-23 March 2021). a Time series of bulk aerosol chemical composition with organic, inorganic (sulfate, nitrate, chloride and ammonium) and equivalent black carbon (eBC) mass concentrations. Shaded pink periods representing the chloride peaks were observed during the late night and early morning periods with the eBC peaks. (b) Mass fraction information throughout the campaign. c, d show the time series of gases (NO₂, NO_X, SO₂, CO, and O₃) together with PM_2.5 mass concentrations. e ALWC along with RH (red square), temperature (color coded), and precipitation information during the campaign. The ALWC peaks often coincide with the chloride and RH peaks, demonstrating that the chemical composition of the aerosol plays an important role in enhancing the ALWC with the support of higher RH values. f, g show other important meteorological data such as solar radiation (SR), planetary boundary layer height (PBLH), along with wind speed and wind direction information. The temperature reaches its highest value just after the PBLH reaches its maximum, with a relatively constant wind direction before the washout episode on 22 March.

Diurnal evolution of ground-level aerosol properties

The diurnal variation in aerosol chemical composition, trace gases, black carbon, and meteorological parameters provides critical insights into the processes governing near-surface haze formation in Delhi (see Fig. 2).

Fig. 2: Diurnal variations of aerosol composition, trace gases, and meteorological parameters.

a Average diurnal variation of the bulk aerosol chemical composition mass fraction (organics, inorganics and equivalent black carbon, eBC). The solid white line with circle markers shows the variation of the mean chloride mass concentration. Chloride peaks were identified throughout the campaign during the late night and early morning periods. b Average diel variation of NO_x (solid brown line), NO₂ (solid light green line), SO₂ (solid red line) and PM_2.5 (solid blue line) observed at ground level. The shaded regions show the standard deviations for each species measured. Most of the measured parameters show the trends where peaks were observed during the morning and late night. c Average diel variation of O₃ (solid dark yellow line), CO (solid grey line), and eBC (solid black line) mass concentrations. d Average diel cycle of the ALWC (dashed black line), together with RH (in red dots) and temperature (color coded). The ALWC trend shows higher values during the early morning, similar to the chloride mass concentration profile.

The mass fraction of major aerosol constituents exhibited a pronounced shift from nighttime to daytime. Organics dominated throughout the day (~40-50%), while inorganics such as nitrate and chloride showed significant nocturnal enhancements. The diurnal cycle showed that the chloride concentrations, in particular, increased from ~2 µg/m³ at 00:00 LT to ~8 µg/m³ by 06:00 LT, coinciding with the peak in organics. This trend suggests an increased contribution likely from biomass burning, waste burning activities, and industrial emissions during the evening and nighttime^3,6. Nitrate and ammonium also exhibited an early-morning peak, showing their condensation into the aerosol phase under humid and cold conditions. As the boundary layer height increased post-sunrise, dilution effects caused a rapid reduction in chloride mass concentration, which dropped below 2 µg/m³ by 10:00 LT.

The NO_x levels showed a persistent nighttime build-up, reaching ~150 ppb around 21:00 LT, driven by vehicular and industrial emissions trapped under a stable boundary layer. NO₂ follows a similar pattern, with concentrations of ~70-80 µg/m³ during the night later decreasing post- sunrise due to photochemical loss and dilution. In contrast, PM_2.5 steadily increased from ~80 µg/m³ at 18:00 LT to ~150 µg/m³ at 08:00 LT, reinforcing the role of fresh primary emissions and secondary aerosol formation during haze development. SO₂ concentrations remained relatively low (~30 µg/m³) but increased slightly during early morning, indicating possible contributions from coal combustion and industrial sources.

The ALWC gradually increased from ~10 µg/m³ at 00:00 LT to higher values (~30 µg/m³) at 08:00 LT, closely following the aerosol chemical composition trend, especially an increase in chloride mass concentrations and RH variations. RH exceeded 80% at night, enhancing water uptake and facilitating aqueous-phase reactions that could have promoted secondary aerosol formation³. As the temperature increased beyond 30 °C after 10:00 LT, RH dropped below 40%, leading to a rapid reduction in ALWC and subsequent haze dissipation. The evolution of eBC and O₃ further indicates contrasting sources and transformation mechanisms. The eBC concentrations remained stable (~20 µg/m³) during the night, demonstrating high intensity continuous combustion sources, but showed a dip after 08:00 LT due to boundary layer expansion. O₃, on the other hand, remained suppressed (<20 µg/m³) throughout the night due to titration by NO but increased sharply post-sunrise, reaching ~100 µg/m³ by 14:00 LT as photochemical activity intensified.

The nighttime accumulation of primary pollutants predominantly drives the near-surface haze formation in Delhi, increased ALWC under high humidity, and secondary aerosol formation via heterogeneous processes²⁴. The rapid dissipation of haze post-sunrise is facilitated by boundary layer expansion, reduced RH, and increased photochemical oxidation. These findings underscore the need for emission control measures targeting nocturnal sources and moisture-driven secondary aerosol processes and their understanding, especially in vertical columns, to mitigate haze formation in Delhi.

Effective sampling strategy for drone-based monitoring using PM-LCS

To ensure effective monitoring of airborne PM values using PM-LCS on-board the drone, two primary challenges must be overcome: (i) the impact of rotor downwash on sampling efficiency and (ii) the influence of humidity on PM-LCS performance. We have designed a drone platform that minimizes both these effects, thereby improving the accuracy of the measurement.

Previous studies^42,60 have demonstrated that integrating a silica-gel-based dehumidifier in front of PM-LCS effectively minimizes humidity-related effects. We have also implemented a similar approach in our setup and observed a significant enhancement in the PM-LCS performance. Figure 3 depicts the correlation coefficient (R²) between the PM-LCS onboard the drone and stationary measurements improved from 0.56 (without a dehumidifier) to 0.80 (with a dehumidifier), demonstrating a ~43% increase in measurement reliability. The mean absolute error in PM_2.5 measurements also decreased by ~35%, confirming the role of dehumidification in stabilizing PM-LCS outputs. Dehumidification also led to a reduction in humidity-induced PM_2.5 overestimation, attributed to hygroscopic growth, ~20 µg/m³. This improvement highlights the importance of preconditioning the sampled air to reduce humidity-induced biases, ensuring that PM-LCS provides robust and reproducible PM_2.5 concentration readings.

Fig. 3: Scatterplots between drone-based PM_2.5 measurements and stationary sensor-measured PM_2.5 values, color coded by RH.

Intercomparison experiments conducted prior to the measurement campaign for minimizing the impacts of humidity on PM-LCS performance on 16 March (a) Drone setup without dehumidifier (b) drone setup with dehumidifier, and, for minimizing the impacts of the rotor downwash on the PM_2.5 mass concentration values on 16 March, morning time (c) without sampling inlet (d) with a sampling inlet. The intercomparison experiment was carried out with the drone hovering with the payload for 16 minutes (16 data points with a 1-minute average of PM_2.5 data).

Drones generate complex airflow patterns due to rotor-induced turbulence, which can result in the under sampling of PM and inaccurate concentration readings^51,54. Because of that, the placement of the PM-LCS and/or inlet must be carefully considered to minimize the effects of the rotor downwash.

To quantify this effect, we conducted experiments with and without a dedicated sampling inlet. Without an inlet, the correlation coefficient between on-board PM-LCS and the stationary instrument was poor (R²=0.35), indicating significant particle losses due to rotor-induced flow perturbations. By incorporating a 36 cm sampling inlet positioned vertically in front of the PM-LCS, we achieved a substantial improvement, with the correlation increasing to R² = 0.78. We have also reported similar comparisons for PM₁ data during 16 March 2021 morning time (same time period for PM_2.5). Specifically, we observed a strong correlation (R² = 0.8) between drone-mounted PM₁ measurements and those obtained from a co-located stationary reference sensor when the sampling inlet and dehumidifier were employed (see Supplementary Fig. 1). In contrast, comparison experiments conducted without these components yielded significantly lower correlations (R² = 0.33 and 0.43, respectively), highlighting the importance of effective sampling strategies in sensing airborne PM. Regarding PM₁₀, we have adopted a cautious approach based on the recommendations of Molina Rueda et al.⁶¹, who advised against including PM₁₀ and coarse-mode (PM_2.5-PM₁₀) concentrations derived from low-cost PM sensors. According to their findings, such sensors are more accurately characterized as being predominantly sensitive to the accumulation mode (PM_0.1-PM₁). Any estimation of PM₁₀ mass concentrations should therefore be regarded as a proxy for the contribution of the accumulation mode to the broader PM₁₀ size fraction. Therefore, while we report PM₁₀ values for completeness, we deliberately refrain from validating or discussing these measurements scientifically, as was rigorously conducted for PM₁ and PM_2.5.

We have also provided the comparisons between drone-based PM_2.5 measurements and stationary sensor-measured PM_2.5 values, under moderate RH (depicting relatively dry conditions, RH <70%, see Supplementary Fig. 2). Comparative analysis of PM-LCS performance indicated a marked improvement in the performance of the drone-mounted PM-LCS under elevated humidity conditions, as shown in Fig. 3. Furthermore, strong correlations were observed between PM_2.5 measurements from the stationary sensor and those obtained from the drone when using a dehumidifier (R² = 0.90) or an inlet (R² = 0.93), compared to significantly lower correlations without the dehumidifier (R² = 0.68) and without the inlet (R² = 0.49), as presented in Supplementary Fig. 2. These results underscore the effectiveness of strategic sensor placement and air intake design in mitigating rotor downwash effects, leading to more reliable airborne PM_2.5 data collection. Integrating a silica-gel dehumidifier and a strategically placed sampling inlet significantly enhances the performance of PM-LCS mounted on drones. These optimizations ensure higher accuracy and reliability in airborne PM_2.5 measurements, addressing key limitations in drone-based air quality monitoring.

Vertical profiles of air pollutants during the measurement campaign

To assess the reliability of PM-LCS performance on-board a drone, we conducted inter-comparison tests before the main measurement campaign (Fig. 3 and Supplementary Fig. 3). After getting reliable outputs from the inter-comparison results, a total of 40 drone flights (8 each day) were performed during morning (three each day), noon (two each day), and afternoon (three each day) between 18–23 March 2021. Minimal variation in PM mass concentrations was observed at noon and in the evening, making it difficult to distinguish between hazy and non-hazy conditions at these times. Therefore, our analysis primarily focuses on distinguishing between hazy and non-hazy mornings. We recorded three vertical profiles each morning (from 06:00 to 08:00 LT) for five mornings, except for 22 March. We excluded the data from 22 March due to a sudden GPS malfunction that led to a drone crash. The exact reasons remain unclear and are beyond the scope of this paper.

Figure 4 illustrates the vertical PM profiles recorded during three hazy morning events (18, 19, and 20 March 2021) and one non-hazy event (23 March 2021) as representative ones. On 18 March, PM concentrations exhibited a notable 60% increase with altitude, reaching ~160 µg/m³ at higher elevations compared to ~100 µg/m³ at the surface. This enhancement was attributed to a temperature inversion (due to the impact of regional emissions and low PBLH) near the surface, which trapped pollutants in the lower boundary layer. Supporting meteorological data indicate high humidity and low temperature conditions at higher altitudes (see Supplementary Fig. 4). A progressive decrease in PM concentrations was observed on 19 March, with values declining from ~110 µg/m³ near the surface to ~90 µg/m³ aloft, representing an 18% reduction. Although overall concentrations were lower compared to 18 March, they remained elevated. On 20 March, a sharp inversion layer was detected around 70 m altitude, leading to a sudden increase in PM_2.5 from ~100 µg/m³ to ~130 µg/m³ within a 20 m vertical span, corresponding to a 30% rise.

Fig. 4: Vertical profiles of particulate matter mass concentrations during hazy and non-hazy mornings measured by PM-LCS on-board the drone.

a Mass concentrations of PM measured using the LCS setup installed on-board the drone on 18 March 2021, a case considered as a severe hazy morning. Here, the solid black line represents PM₁ mass concentration, the solid red line represents PM_2.5 mass concentration, and the solid blue line represents PM₁₀ mass concentration. The error bars for each respective measured concentration data showed the standard deviation values. b particulate matter values vertically measured by drone setup during a hazy morning on 19 March 2021. c particulate matter values vertically measured by drone setup during a hazy morning on 20 March 2021 d particulate matter values vertically measured by drone setup during a non-hazy morning on 23 March 2021.

In contrast, a major rain event prior to 23 March led to a significant reduction in PM levels, resulting in a non-hazy morning. PM_2.5 concentrations at the ground level dropped to ~40 µg/m³, marking a ~60% decrease compared to 18 March. The concentration remained relatively stable with altitude, maintaining a nearly constant profile up to 100 m above ground level (a.g.l.), indicating the absence of strong inversion effects. We have also provided the vertical variability in PM mass concentrations for noon and evening periods (see Supplementary Fig. 5). The vertical profiles showed minimal deviations in PM values and were found to closely resemble the non-hazy morning episode. We have also observed that the boundary layer height (BLH) reached its minimum around 04:00 local time (LT), reaching a mean value of approximately 80 m and an associated ground-level wind speed of 0.5 m/s. Sunrise typically occurred between 06:00 and 08:00 LT. BLH values remained low (80-100 m) during the early morning drone-based measurements conducted between 06:00 and 07:00 LT, coinciding with relatively weak ground-level winds (~0.8 m/s). Note that wind speed data were only available at the surface. On the morning of 20 March 2021, during a pronounced haze episode, a shallow boundary layer (80 m) combined with low wind conditions (0.8 m/s) and elevated relative humidity (~70%) contributed to the accumulation of pollutants. This resulted in elevated PM concentrations near 80 m altitude, as depicted in Fig. 4. On previously discussed hazy episodes, such as on March 18, 2021, the boundary layer was observed to be approximately 120 m (Fig. 1), with notable concentrations of PM detected at 100 m (Fig. 4). Conversely, on the non-hazy morning of March 23, 2021, the BLH increased to around 500 m leading to turbulent mixing, accompanied by slightly higher wind speeds of 1.5 m/s and subsequent rainfall, resulting in significantly lower concentrations of PM compared to hazy mornings. During noon (12:00-13:00 LT) and evening profiles (17:00-18:00 LT), mean BLHs reached 1.5 km leading to turbulent mixing, with wind speeds peaking at 2 m/s. Additionally, lower humidity levels (~ 40-50%) and higher temperatures (30-35°C) contributed to notable reductions in PM concentrations during drone measurements. These findings highlight the pronounced impact of meteorological conditions on vertical PM_2.5 distributions, particularly under inversion-driven haze episodes. Additionally, Supplementary Fig. 6 provides a detailed visualization of the directional distributions of various aerosol components using polar plots. These plots offer insights into the spatial patterns of wind-pollution interaction.

Here, we also examine the vertical variability of the PM₁/PM_2.5 mass ratio in hazy and non-hazy conditions at different times of the day (Fig. 5). Building on the approach proposed by Hodoli et al.⁶², who used the PM_2.5/PM₁₀ ratio derived from micro-sensor data as a proxy for source classification tied to meteorological drivers, we extend this methodology to PM₁/PM_2.5 ratios. We use this ratio as a diagnostic proxy to infer aerosols’ composition and their sources’ characteristics. On the hazy morning of 18 March 2021, we observed a distinct decrease in the PM₁/PM_2.5 ratio with altitude, starting with a value of ~0.56 and reaching a value of ~0.47 at 100 m (see Fig. 5a). This indicates a dominant contribution of more than 50% from larger particulate fractions. This is likely due to secondary inorganic aerosol species exhibiting higher hygroscopicity (κ) values at higher altitudes (RH > 70%) than at ground level (RH < 70%) more consistent with the findings from Sun et al.⁶³. This behavior is consistent with enhanced water uptake and growth of hygroscopic particles, suggesting secondary aerosol formation aloft. A similar pattern emerged on 20 March 2021, with a reduced PM₁/PM_2.5 ratio of ~0.5 noted at higher elevations during the morning. This period was characterized by temperature inversions and a shallow boundary layer of ~80 m (see Fig. 1), further implicating secondary haze formation processes. The 19 March profile represents an intermediate haze condition between the severe 18 March and the more confined 20 March events. During the morning, the PM₁/PM_2.5 ratio remained nearly uniform with altitude (~0.65-0.68) under relatively high humidity (60-70 %), suggesting a partially mixed layer without strong vertical stratification. Unlike the pronounced layering observed on 18 March, the morning ratio profile showed only a slight decrease with height, implying moderate hygroscopic growth and limited secondary aerosol formation aloft. By noon and evening, the ratios exhibited a gradual increase toward ~0.7-0.75, coinciding with slightly lower RH values (40-55 %), consistent with the dissipation of morning stability and partial dilution of accumulated fine particles. This pattern indicates a transitional stage in the haze cycle, residual humidity and pollutant build-up persisted from the prior night’s inversion, but increasing solar heating began to erode stability, leading to partial vertical mixing. The 19 March case thus bridges the strongly stratified haze of 18 March and the shallow, near-surface inversion of 20 March, demonstrating how the progression of boundary-layer recovery controls the vertical redistribution and microphysical evolution of aerosols through the day. In contrast, on the non-hazy morning of 23 March 2021, the PM₁/PM_2.5 ratio remained high (0.62-0.67), which indicates the dominance of primary emissions, most likely from vehicular and industrial sources. These findings are consistent with a previous study reporting PM₁/PM_2.5 ratios decreased markedly with RH > 60% during periods dominated by SIA (>50%), unlike primary organic aerosol (POA) from traffic or coal combustion, which exhibited limited sensitivity to RH⁶⁴. We demonstrate that ratios greater than 0.5 are indicative of fine-mode aerosols associated with primary emissions from combustion and semi-volatile secondary components (e.g., organics). In contrast, lower ratios reflect the dominance of the coarse mode, high hygroscopic particles (nitrate, sulfate, ammonia, and chloride) that shift the effective diameter closer to that of PM_2.5^64,65. During other time intervals (e.g., noon and evening) across all measurement days, PM₁/PM_2.5 ratios were consistently higher (above 0.6, reaching up to 0.8 at noon), suggesting fresh primary emissions with lower particle diameters and likely higher organic content. These findings highlight the usefulness of PM₁/PM_2.5 ratios as an effective proxy for distinguishing between primary and secondary aerosol sources and their transformation processes, particularly when combined with RH data. This approach provides a valuable means of inferring particle hygroscopic growth and associated haze formation dynamics in complex urban environments where direct chemical composition measurements may be limited. While the mechanisms described here are derived from a limited set of representative cases, they are consistent with established theory and observations over the region. Future multi-seasonal vertical profiling and coupled modeling will be essential to confirm the persistence and variability of these mechanisms under different meteorological regimes.

Fig. 5: Variations of PM₁/PM_2.5 ratios and RH from LCS at different times of day and altitudes during hazy and non-hazy conditions.

a depicts the PM₁/PM_2.5 ratio calculated from the PM-LCS data on-board the drone on 18 March 2021, a case considered as a severe hazy morning along with noon and evening PM ratio profiles. The ratio at different times of the day is shown at different RH values. b vertical PM ratio values on 19 March 2021. c vertical PM ratio values on 20 March 2021 (d) vertical PM ratio values on 23 March 2021.

Model validation with the vertically resolved data

Here, we compared the drone-measured PM_2.5 values with WRF-Chem simulated PM_2.5 values. Based on the drone-based vertical profiling of PM_2.5 concentrations, the WRF-Chem model consistently underestimated observed values across all altitudes and days, with varying degrees of deviation³⁹. The observational data revealed significant discrepancies in WRF-Chem’s predictions of RH and PM_2.5 concentrations during morning hours, with errors markedly amplified under hazy conditions. Figure 6a shows that during hazy mornings, WRF consistently underestimated RH across the entire vertical profile (0-100 m). For example, on March 18, 2021, WRF predicted surface-level RH values between 32-34%, while drone measurements ranged from 62-74%, corresponding to relative differences exceeding 50%. This dry bias extends through higher altitudes, reaching a maximum relative difference of 76.6% at 100 m on the same day. In contrast, during non-hazy mornings (e.g., March 23, 2021), WRF predictions of RH are much closer to drone observations, with differences generally below 10% across the vertical profiles. The smaller discrepancy under non-hazy conditions highlights that WRF-Chem’s RH underestimation is particularly severe during hazy, polluted mornings. Next, the mean percentage difference in RH between WRF-Chem simulations and drone observations remains below 10% across altitudes during non-hazy mornings. In contrast, during hazy mornings, this difference exceeds 55% and reaches over 70% at certain altitudes (Fig. 6b), indicating that approximate RH underestimation by the model is more than five times greater under hazy conditions. Furthermore, to assess a qualitative indication of the linkage between RH bias and PM_2.5 underestimation, the relationship between RH bias and underestimated PM_2.5 percentage were analyzed in tandem (Fig. 6c). A statistically significant relationship was observed between larger negative RH biases coinciding with greater underestimation of PM_2.5 concentrations by WRF-Chem during hazy mornings. The mean RH dry bias during hazy conditions is approximately −30.1 ± 8.3%, compared to −3.0 % during non-hazy periods, coinciding with PM_2.5 underestimation of −52.6 ± 5.5% in hazy vs. −10.8 ± 1.2% in non-hazy mornings (Fig. 6c), respectively. A strong positive correlation (Spearman’s r = 0.89, p-value< 0.001) between RH bias and PM_2.5 underestimation further suggests that the model’s dry bias likely impairs its ability to simulate aerosol hygroscopic growth, thereby contributing to systematic underprediction of particulate matter under humid, polluted conditions.

Fig. 6: Influence of vertical relative humidity on PM_2.5 prediction during hazy and non-hazy conditions: comparison between drone measurements and WRF-Chem simulations.

a Vertical profiles of drone-measured and WRF-simulated RH values during hazy and non-hazy mornings. b Mean % RH difference with altitude during hazy and non-hazy mornings. c RH bias and subsequent uncertainty percentage in PM_2.5 prediction by WRF-Chem when compared with the drone-measured PM_2.5 values.

While, without aerosol composition data (e.g., hygroscopicity parameter κ), we cannot definitively partition errors between RH and other factors on PM quantification. But to qualitatively see if the non-hazy morning period’s lower RH bias and PM underestimation is consistent, we quantify the same parameters for the noon and evening periods (Supplementary Fig. 7). While a clear classification of hazy versus non-hazy conditions is not available for the noon and evening periods, the observed biases in relative humidity (RH) and PM_2.5 during these times are notably smaller than those observed during hazy mornings and closely resemble the non-hazy morning conditions. Specifically, the mean RH bias during noon and evening ranges from approximately −3.8% to −7.5%, with a combined average of −5.7%, comparable to the −3.0% bias recorded during non-hazy mornings. Similarly, PM_2.5 underestimation by WRF-Chem during noon and evening varies between −12.9% and −22.0%, averaging around −17.4%, again aligning more closely with the −10.8% underestimation observed on non-hazy mornings than the substantially larger −52.6% underestimation during hazy mornings (see Supplementary Fig. 8).

Recent observational and modeling studies have highlighted the critical role of relative humidity (RH) in influencing aerosol formation and model accuracy under high-humidity conditions, particularly during fog and haze episodes in severely polluted urban environments such as Delhi. Acharja et al.⁶⁶ reported a significant enhancement in secondary inorganic aerosol (SIA) concentrations during fog events, with mass concentrations increasing notably from non-fog to fog hours. These results underscore the importance of fog-driven aqueous-phase chemistry in urban aerosol evolution, suggesting that conventional model representations, such as in WRF-Chem, often fail to capture these processes adequately.

Jat et al.⁶⁷ further demonstrated the sensitivity of model performance to fog events by excluding fog-affected days from model evaluation. This markedly improved the model’s ability to predict PM_2.5, indicating that the current parameterizations inadequately represent fog-related microphysics and chemistry. Since fog and haze often co-occur under high-humidity, stagnant conditions in winter, these findings have broader implications beyond fog episodes alone. Hazy conditions, although not always associated with saturation and droplet formation, also likely observe enhanced heterogeneous and multiphase chemical pathways, especially when RH is elevated.

The planetary boundary layer height (PBLH) also plays a central role in modulating PM_2.5 levels, with a stronger influence than other meteorological parameters^68,69. To isolate chemical and emission-related uncertainties from boundary-layer effects, Jat et al.⁶⁷ further assessed PM_2.5 predictions on days when modeled and observed PBLH values were closely aligned (within ±50 m). The persistent underprediction of PM_2.5 on these days’ points toward deficiencies in chemical mechanisms and emission inventories, specifically, the exclusion of anthropogenic chloride sources and the incomplete representation of ammonia–HCl reactions^70,71. Our findings of a persistent dry bias and PM_2.5 underestimation during hazy mornings point to a critical limitation in the model’s physical representation. While we attribute the PM bias to the RH-driven underprediction of hygroscopic growth, the RH bias itself likely stems from a deeper model deficiency. Recent work over the IGP has highlighted a positive feedback loop wherein the hygroscopic growth of aerosols increases aerosol optical depth, leading to surface cooling, moistening of the boundary layer, and enhanced stability, which further traps pollutants and amplifies haze⁷². Our WRF-Chem configuration, which does not include aqueous aerosol chemistry, hygroscopic growth, and their interaction with meteorology, could be the reason of it being incapable of simulating this self-amplifying process. Consequently, the model fails to generate the observed levels of near-surface humidity and boundary layer stability, leading to the concurrent dry bias and underestimation of PM_2.5. This omission, combined with potential inaccuracies in the ABL parameterization under very stable conditions⁶⁷, explains why the model performs poorly during haze episodes despite a reasonable bulk agreement on PBLH (±50 m), a metric that masks the critical internal stratification and humidity profiles we observed. In addition, emissions from different sources, including waste-burning, agricultural residue burning, etc, are largely absent from current inventories, particularly during winter heating in peri-urban and rural regions⁷³. Observational studies have shown that chloride-containing aerosols significantly contribute to Delhi’s PM_2.5 burden and secondary aerosol formation. The lack of chloride emissions in the model leads to systematic biases in both surface PM_2.5 concentrations and aerosol optical depth (AOD), especially across northern India⁶⁷. Pithani et al.⁷⁴ also revealed a consistent underestimation of near-surface RH (~15-20%) by WRF-Chem during dense fog events. Similarly, discrepancies in humidity prediction using the WRF-Chem model over an intense haze event in Delhi during November 2016 were also demonstrated in Agarwal et al.⁷⁵. This RH bias, often arising from temperature errors of 1-2 K, prevents the model from reaching saturation, thereby inhibiting fog formation and the associated aqueous-phase and multiphase reactions. These conditions suppress fog development and reduce the extent of aerosol hygroscopic growth and secondary aerosol production, equally relevant mechanisms during hazy but unsaturated high-RH episodes^67,74. Thus, the limitations observed in fog modeling under high RH conditions in Delhi also apply to haze events, which are similarly characterized by elevated humidity, suppressed boundary layer mixing, and enhanced pollutant trapping. The underprediction of PM_2.5 during fog and haze episodes reflects a broader need to improve RH representation, fog/haze process parameterization, and the inclusion of missing emissions and heterogeneous chemistry in chemical transport models like WRF-Chem. Improved simulation of RH and associated chemical processes will be essential for reducing model biases, particularly during winter when persistent fog and haze events contribute substantially to poor air quality.

These findings suggest that the more stable and humid boundary layer conditions characteristic of hazy mornings exacerbate WRF-Chem’s dry bias and consequent underprediction of PM_2.5, likely due to misrepresentation of aerosol hygroscopic growth. In contrast, the reduced RH and PM_2.5 biases during noon and evening imply that the model performs better under less stagnant and more mixed atmospheric conditions.

Based on the Supplementary Fig. 5, the WRF-Chem model and the drone-based PM_2.5 values comparisons were provided. These findings emphasize the need for improved model parameterizations, particularly in severe haze scenarios, where discrepancies are highest. Incorporating high-resolution vertical measurements into WRF-Chem validation is crucial for enhancing urban air quality forecasting, as previous studies have also highlighted similar challenges in simulating vertical aerosol distributions in the lower boundary layer³⁹.

Discussion

The drone-mounted PM-LCS employed in this study was used to monitor vertical profiles of PM. Operationally, the PM-LCS mounted on-board the drone successfully reduced humidity biases (20 µg/m³) and rotor interference (R² improvement from 0.35 to 0.78). This provides an effective sampling protocol for drone-mounted PM-LCS use in air pollution monitoring. Additionally, the measurements reveal previously unrecognized complexities in Delhi’s haze dynamics. We provided the PM₁/PM_2.5 mass ratio as a proxy indicator of the aerosol chemical composition, which revealed the near-surface haze formation that may not be completely resolved by ground-level monitoring and coarse-resolution modeling. Our findings build upon previous research into aerosol thermodynamics and secondary formation under high RH conditions by directly observing vertical structures during haze episodes. For example, on the hazy morning of 18 March 2021, we observed a significant decrease in the PM₁/PM_2.5 ratio with altitude (from ~0.56 at ground level to ~0.47 at 100 meters), indicating a greater contribution from larger secondary particles at higher elevations. This vertical shift is consistent with enhanced aqueous-phase processing under elevated RH (>70%) aloft, likely involving hygroscopic SIA species such as ammonium nitrate and ammonium chloride. These observations align with those of Sun et al.⁶³, who demonstrated the RH sensitivity of SIA-driven haze events in China. In contrast, the consistently high PM₁/PM_2.5 ratio observed during the non-hazy morning of 23 March 2021 (0.62-0.67) implies a dominant role for primary emissions, likely from vehicular or industrial activity, under a less humid boundary layer regime.

We also explain the probable yet critical role of aerosol phase state reactivity supported by ground-based measurements. The elevated nocturnal and early morning chloride during episodes (20-30% of PM_2.5) and nitrate (10-15%) concentrations, coupled with relative humidity exceeding 80%, suggest that the aerosol particles could transition into a liquid or semi-solid phase (multiphase chemistry) during the haze episodes. This phase shift could enhance heterogeneous reactions, such as chlorine-driven oxidation of organic precursors, likely accelerating secondary aerosol formation, a process generally not much considered in models that assume solid-phase equilibria. This in turn affects other pollutants like BC, leading to their extended lifetime, change in column mass concentrations⁷⁶ as well as non-linear impact on direct and indirect forcing⁷⁷. In addition, morning haze peaks (05:00-08:00 LT) coincide with suppressed boundary layer heights (<200 m) typically observed in winter and post-winter months in Delhi and stagnant conditions that favor condensation of semi-volatile species (e.g., NH₄Cl, NH₄NO₃). In contrast, daytime boundary layer expansion disperses aerosols. Still, it oxidizes precursors (e.g., NO reacting to HNO₃), resetting the system for nighttime accumulation, a cycle that models fail to resolve without diurnally varying chemistry schemes.

This study also provides compelling evidence that the WRF-Chem model fails to accurately simulate the key processes that govern haze formation, particularly in high-humidity conditions. Model-predicted RH values during hazy mornings were severely underestimated, by over 50% in some cases, relative to drone observations (e.g., 32-34% vs. 62-74% on 18 March 2021), particularly near the surface and throughout the 0-100 m layer. This dry bias was consistently linked to underpredictions of PM_2.5 mass concentrations by over 50%, reinforcing failure of the model to capture RH-driven secondary formation. The strong positive correlation (Spearman’s r = 0.89, p < 0.001) between RH bias and PM_2.5 underestimation confirms that this limitation stems from the inadequacies of the WRF-Chem model in simulating the uptake of water by aerosols and the related multiphase chemistry under humid, polluted conditions. This finding is consistent with that of Acharja et al.⁶⁶, Jat et al.⁶⁷ and Agarwal et al.⁷⁵, who emphasized that model performance deteriorates during fog or haze events in Delhi. Furthermore, the inability of the model to predict aerosol phase transitions, such as the shift to a liquid or semi-solid state at RH > 80%, hinders the accurate prediction of heterogeneous processes, including chlorine-driven oxidation and semi-volatile condensation. These processes influence the burden and optical and radiative properties of PM_2.5. Overall, these dynamics and distribution likely explain the systematic mean underestimation of PM_2.5 mass concentrations predicted by WRF-Chem during hazy mornings, as current parameterizations overlook dynamic mixing and size variations (which impact mass), hygroscopic growth and aqueous-phase chemistry.

The differing spatial and temporal resolutions between the WRF-Chem model (3 km grid, hourly outputs) and the point-based drone measurements represent a recognized challenge. This mismatch complicates direct comparison and may reduce confidence in evaluating model accuracy, particularly in the highly heterogeneous and rapidly evolving lower atmospheric boundary layer. Such discrepancies may lead to apparent inconsistencies in pollutant concentrations and vertical structure, complicating the identification of model biases. Nonetheless, prior studies have demonstrated that vertical drone observations remain valuable for diagnosing model performance despite these challenges. For instance, Liu et al.³⁹ demonstrated that vertical drone observations could reveal key model deficiencies -especially under stable nocturnal boundary layers, despite using a coarser (~4 km) WRF-Chem configuration. Similarly, Zhu et al.⁷⁸ validated the Community Multiscale Air Quality Modeling System (CMAQ) outputs via drone-based ozone profiles, effectively diagnosing planetary boundary layer height and emission uncertainties despite spatial resolution gaps. Lee and Kwak⁷⁹ combined drone measurements with WRF-CFD simulations to characterize near-road pollution, illustrating that high-resolution vertical profiles yield meaningful insights despite limited spatial representativeness. This study also highlights the effectiveness of drone-based vertical profiling in diagnosing model deficiencies, even amidst the potential spatial-temporal mismatches. Although drone data are point-based and temporally limited (40 vertical profiles), they reveal the intricate structures of PM_2.5 and RH gradients that are crucial for model enhancement. This is consistent with previous studies by Liu et al.³⁹, Zhu et al.⁷⁸ and Lee and Kwak⁷⁹, which demonstrated the usefulness of vertical drone measurements in validating the outputs of coarse-grid models and their representations of the boundary layer.

The present study highlights several critical gaps in understanding haze formation in Delhi. While the drone-mounted PM-LCS provided vertical PM profiles and proxy indicators (e.g., PM₁/PM_2.5 ratios) to reveal near-surface and elevated aerosol dynamics, key processes remain unresolved by both low-cost sensors and current models. These include vertical variations in ultrafine particle growth, aerosol chemical composition (secondary inorganic and organic species), phase state transitions, particle mixing state, and hygroscopicity, all of which strongly influence nocturnal and early morning haze initiation. Furthermore, models used during the study such as WRF-Chem fail to capture high-humidity conditions, multiphase chemistry, and boundary layer dynamics, leading to significant underestimation of RH and PM_2.5 mass, as well as misrepresentation of secondary aerosol formation and optical properties. Future studies should complement such measurements with high-resolution, multi-platform observations to capture ultrafine particle dynamics, nocturnal growth at higher altitudes, secondary organic aerosol formation, and inorganic composition aloft. Integration of advanced miniature aerosol instruments with ground-based supersites and remote sensing will provide detailed insights into aerosol evolution and radiative impacts, enabling more robust prediction and mitigation of severe haze events.

Nevertheless, this work represents a significant advancement in integrating low-cost PM sensors with advanced model diagnostics, demonstrating their potential to bridge observational gaps in the lower boundary layer. By directly quantifying the interactions between RH and PM, as well as model biases, under real-world haze conditions, our findings pave the way for next-generation air quality models that incorporate aerosol chemistry and dynamic boundary-layer coupling. These innovations are critical not only for improving forecasts and public health interventions in megacities such as Delhi, but also for informing global strategies aimed at mitigating air pollution and its climate feedbacks in rapidly urbanizing regions.

Methods

Measurement site

The aerosol sampling site at Indian Institute of Technology (IIT Delhi) (28.5457° N, 77.1928° E, 230 m a.s.l.) is situated in South Delhi, amidst residential, educational, and commercial buildings, exemplifying a densely populated urban environment²⁶. A major arterial road, approximately 150 m away, experiences heavy traffic, with sources emitting significant quantities of PM, black carbon (BC), and NO_x^6,7,25. Previous studies indicated that vehicular emissions contribute up to 40-50% of PM_2.5 during peak hours in Delhi, with equivalent BC (eBC) concentrations reaching 10-20 µg/m³ near traffic corridors²⁶. Additional sources of emissions include residential cooking and heating, waste burning, as well as industrial activities. Gunthe et al³. highlighted the role of ammonium chloride (NH₄Cl), linked to urban hydrochloric acid (HCl) emissions, in enhancing aerosol water uptake, a process likely significant at this site due to waste burning and e-waste industrial activities. In addition to the emission sources, Delhi faces its own unique meteorological challenges. The wintertime boundary layer height over Delhi, even during the late winter months such as February (~50-200 m; S. Raj et al.²⁰) leads to pollutant accumulation, while dilution effects dominate in summer (~1000-1500 m). Given the preponderance of local sources in determining PM_2.5 concentrations, often exceeding 150 µg/m³ in winter^3,5, the IIT Delhi site offers a suitable location for studying aerosol characteristics. These characteristics include chemical composition, hygroscopicity, and meteorological interactions, particularly in a high-emission urban setting. The sampling locations are illustrated in Supplementary Fig. 9, which shows the Delhi area, the specific sampling location at the Indian Institute of Technology (IIT) Delhi, the flight path of the drone and the drone’s ascent and descent patterns.

Flight platform with a modified sampling system

A drone platform was custom-designed by BotLab Dynamics Ltd., India (a drone start-up company, https://botlabdynamics.com/, last accessed on 17^th November 2025) for air quality monitoring purposes. The drone platform used during the measurement campaign is a multi-rotor system with four rotor arms. Lithium-polymer (LiPo) batteries attached to the drone platform provide an endurance of 15-20 min, depending on the payload specifications used during the flights. The platform, including the batteries and the scientific payload, weighs about 7.5 kg with a payload capacity of around 1.2 kg. The ascending speed of the drone was maintained at 1.5 m/s. A PixHawk version 2.4 flight controller was used for communication purposes. The drone, including its payload, maintained a stationary hover for 150 seconds or 2.5 minutes at each 20-m altitude increment.

To improve the performance and reliability of the PM-LCS system, significant attention was devoted to enhancing the enclosure’s airtightness and minimizing unintentional air leaks. In particular, we identified and addressed leakage points near the data output cables- an area prone to small but critical leaks that could otherwise lead to flow reductions, altering internal pressure dynamics, and reducing the sampling efficiency of the PM-LCS⁸⁰. These leaks were then sealed using a combination of high-grade silicone-based sealants and custom-fitted grommets to ensure mechanical flexibility while maintaining a robust seal. Leak testing was later performed using smoke visualization to ensure complete sealing under operational flow conditions.

In an effort to enhance sample flow stability, we have replaced the low-power regular fan of PM-LCS with an external rotary vane pump of 2.7 lpm (Metzger Technik GmbH, Germany). Our approach of replacing a regular fan with an external pump and an orifice aligns with the approach proposed by Bezantakos et al.⁸¹, which demonstrated that the integration of an external pump into a low-cost optical particle counter (OPC) system enhances flow stability and ensures a consistent particle throughput without compromising instrument performance. The modification was done to minimize the particle losses due to the lower suction rate of the usual fan present in the PM-LCS. This helps to collect particles more systematically when a sampling inlet and dehumidifier (or drier) are placed upstream of the PM-LCS.

An aerosol vertical sampling inlet 36 cm in height (height selection after performing the in-flight inter-comparison experiments) was used for effective sampling to minimize particle losses due to rotor interferences. A 3-D printed connector was explicitly designed for this type of PM-LCS to connect the vertical sampling inlet to the pump. The sampling inlet was connected to a splitter, which connected both the PM-LCS and AE-51 micro-Aethalometer instruments. A custom-built silica-gel dehumidifier was placed between the splitter and the sampling tube. The main goal of this type of dehumidifier in this configuration is to achieve a reduction of relative humidity (RH) to around 40% for at least the duration of the drone’s flight endurance, which can be described as a dry flow where particle water uptake is negligible⁸². Such a dehumidifier is, therefore, intended for short-term measurements on a drone platform instead of longer-term use on a stationary sampling station. The drying duration was effective for a period longer than the flight endurance of the drone, which is at least 30 minutes long⁴². An overview of the sampling location and the drone-based sensor configuration is shown in Fig. 7, while a visual representation of one of the drone flights during a hazy morning event at the IIT Delhi sampling location is provided in Supplementary Fig. 10, which captures the drone in flight against the backdrop of the hazy conditions.

Fig. 7: Overview of the study area and drone-based PM-LCS measurement system.

a Map of Delhi showing the sampling site along with industrial facilities located in the region and nearby areas. A background map showing various industrial locations was adopted from Rai et al.¹¹⁵ (which is an open access article distributed under the terms of the Creative Commons CC-BY license 4.0), also shown in our article (Mandariya et al.⁵). b Picture of the drone and its payload consisting of PM-LCS, AE-51 micro-Aethalometer, and meteorological sensors. An aerosol sampling inlet and the silica-gel dehumidifier (dryer) are placed before the payload. In this paper, we primarily use data from the PM-LCS on-board the drone. c Picture of the drone carrying the payload during the measurement flight.

In-flight instrumentation

A cost-effective air quality monitoring device, referred to as PM-LCS (Fig. 7b), was used⁸³. The price of PM-LCS is approximately USD 200 and is designed for real-time measurement of PM and meteorological parameters. The device dimensions are 20.3 cm in length, 12.1 cm in width, and 7.6 cm in height, with a total weight of 500 g. The PM-LCS monitor has a Plantower PMS7003 sensor (USD ~25; dimensions: 4.8 cm × 3.7 cm × 1.2 cm), which uses light scattering from particles illuminated by a laser within the sensing chamber to determine PM₁, PM_2.5 and PM₁₀ concentrations. Air is drawn through the device by a fan (modified in our study; instead of a fan we used a pump) and the scattered light, which is detected by a photodiode, is processed using proprietary algorithms to estimate particle mass concentration⁸⁴. Despite its affordability and high temporal resolution, uncertainties in the measurements arise from the following: assumptions of spherical particle shape and a constant refractive index; hygroscopic particle growth under high humidity; variations in aerosol composition; and factory calibration biases. Further errors may result from degradation of the optical components and reduced counting efficiency for particles outside the sensor’s optimal detection range. The measurement principle and sensor configuration are analogous to other models in the Plantower series, such as PMS3003 and PMS5003, and have been extensively characterized in previous research^85,86,87. Additionally, the device incorporates an Adafruit DHT22 sensor for temperature and RH measurements alongside a Quectel L80 GPS module for accurate real-time geolocation⁸³. Data acquisition follows a structured protocol wherein PM and meteorological parameters are recorded every 3s via a serial TTL interface, aggregated at 1-minute intervals, and subsequently transmitted via a Quectel M66 GPRS module. In cases of connectivity failure, data storage is maintained locally on a microSD card. A fail-safe 3.7V, 2600 mAh rechargeable Li-ion battery ensures uninterrupted operation. Under standard operating conditions, the battery provides up to 10 hours of power at a 1-minute interval. The DHT22 sensor ensures reliable meteorological measurements, complementing the high-quality PM data obtained from the PMS7003 sensor⁸³. More details about the laboratory calibration of a similar sensor are provided elsewhere^83,87. This PM-LCS was modified (as mentioned above in “Flight platform with modified a sampling system“) to make it suitable for drone-based monitoring. During the campaign, a meteorological sensor (SHT31, Sensirion, https://sensirion.com/media/documents/0800F291/671A0E26/SE_Flyer_Humidity_Sensors_EN_web_240616.pdf) was used on-board the drone to measure meteorological parameters such as temperature and humidity. The SHT31 sensor provided temperature and RH values using bipolar junction transistors (BJTs) and a polymer-based capacitive measurement principle, respectively. Minor uncertainties may arise from sensor drift, surface contamination, and response lag under rapidly varying environmental conditions. Additionally, the drone platform was equipped with the AE-51 micro Aethalometer for equivalent black carbon (eBC) concentration estimation (reported in a companion paper), a device weighing approx. 280 g, 11.7 cm in length, 6.6 cm in width, and 3.8 cm in height. The AE-51 stores data in its built-in storage, which can be downloaded later using the provided software. The specifics of the field campaign, including measurement periods and instruments used, are detailed in Supplementary Table 1.

Ground-based instrumentation

An Environmental Beta Attenuation Monitor (E-BAM, model: 1022, manufactured by Met One, USA) was used to evaluate the performance of the PM-LCS, which is a high-precision research instrument⁸⁸. The E-BAM was positioned alongside the PM-LCS on the rooftop of the library building at IIT Delhi (the working conditions were controlled at under 26 °C and 40 % RH) for the duration of four days. This relatively short time frame was selected due to constraints related to the COVID-19 pandemic, which limited ground-based inter-comparisons to four days of a two-week campaign period. The remaining time was allocated for in-flight inter-comparisons and the drone-based flight campaign employing the PM-LCS. The E-BAM measures the attenuation of beta radiation to detect concentrations of PM_2.5 in the sampled air. Data from the E-BAM were recorded at 15-minute intervals during this collocation. The hourly averaged PM_2.5 concentrations from both the sensors and E-BAM were then utilized for further analysis, including conducting statistical tests such as R² tests via linear regression to evaluate the agreement between the sensor-recorded data and the E-BAM measurements. The results indicated that the PM-LCS performed well against the reference instrument, considering the high coefficient of correlation (R²=0.95) (See supplementary Fig. 3). These performance statistics are corroborated by a recent study showing the robust performance of this PM-LCS⁸⁸. These experiments were performed before the in-flight experiments and the actual field campaign.

A quadrupole aerosol chemical speciation monitor (Q-ACSM; Aerodyne Inc., MA, USA) was deployed on the second floor of the building at IIT Delhi, about ~15 m above the ground level, to measure the composition of non-refractory PM_2.5 with vaporization temperatures below 600 °C at 10 min time resolution. These particles include organics (Org), sulfate (SO₄²⁻), nitrate (NO₃⁻), ammonium (NH₄⁺), and chloride (Cl⁻). A stainless-steel tube with a ¼ inch outer diameter and a flow rate of 1.7 lpm was utilized to collect ambient air samples in a well-maintained, temperature-controlled room. From this, a subsample of 0.1 lpm was directed into the ACSM. At the start of the sampling process, a URG cyclone with a 2.5 μm cut-off size was incorporated to filter out larger particles. To maintain sample air RH below 40%, an automatic aerosol Nafion dryer was positioned at the ACSM inlet. The sampling inlet was installed on the roof, directly above the instrument, ensuring a residence time of approx. 5s in the sampling line²⁴. The ACSM was set to operate with a 10-m resolution during the measurement campaign, and the data were later averaged to an hourly interval. Data gaps at the ground station on March 13-14, 2021 resulted from the scheduled calibration of the Q-ACSM instrument. Despite that, we could still observe the chloride trends in the latter half of the measurement period. We have also excluded data for other species for these two days (when ACSM data was missing) to maintain symmetry with the chemical composition information.

The continuous real-time eBC measurement was also performed with the help of a seven-channel Magee® AE-33 aethalometer²⁴. The instrument was used to measure eBC mass concentration at 1-min intervals, and operated at a flow rate of 2 lpm with a PM_2.5 inlet. Auxiliary measurements, including details of the AE-51 micro Aethalometer used for eBC measurements, are described in the Supplementary Note.

External air quality and meteorological data

We have also reported hourly ambient surface air pollutant concentrations and meteorological variables near the IIT Delhi. Ground-based pollutant concentrations, including particulate matter (PM_2.5; μg/m³), nitrogen dioxide (NO₂; μg/m³), sulfur dioxide (SO₂; μg/m³), ozone (O₃; μg/m³), nitrogen oxides (NOₓ; ppb), and carbon monoxide (CO; mg/m³), were sourced from the Central Pollution Control Board (CPCB), Continuous Ambient Air Quality Monitoring (CAAQM) network. While the study area is centered at IIT Delhi, the absence of a CPCB monitoring station at this location required the use of data from the nearest available station in R.K. Puram (28.563°N; 77.186°E), located approximately 5 km west-southwest (W-SW) of IIT Delhi. This station was selected to ensure spatial representativeness, given its proximity and shared regional emission and meteorological characteristics with the study site. Furthermore, meteorological variables, including ambient surface temperature (°C), total precipitation (mm), wind speed (m/s), and wind direction (degrees), were also obtained from the R.K. Puram CPCB dataset. CPCB datasets have been widely used in previous research, investigating regional air pollution over Delhi and other parts of India^89,90,91,92, as well as a reliable benchmark data source to validate against low-cost sensors and remote sensing-based pilot studies^93,94. The CPCB datasets can be sourced from their online portal (URL: https://app.cpcbccr.com/ccr/#/caaqm-dashboard-all/caaqm-landing).

Additional critical meteorological parameters were extracted from the ERA5 reanalysis dataset to complement the ground observations. ERA5 is the fifth iteration of the European Centre for Medium-Range Weather Forecasts (ECMWF) global atmospheric reanalysis, offering comprehensive weather and climate data dating back to 1940. The datasets provide predictions at a 31-km horizontal resolution and incorporate 137 vertical levels to model atmospheric processes, generated using a 10-member ensemble approach^95,96,97. Both the ERA-5 meteorological parameters, boundary layer height (BLH; m) and surface net solar radiation (SR; W/m²) used here are extracted at a spatial-temporal resolution of 0.25° × 0.25°; hourly, with the grid encompassing IIT Delhi selected for analysis. The BLH data in ERA5 are derived from the bulk Richardson number method⁹⁸. While for solar radiation, ERA-5 implements the Rapid Radiative Transfer Model for General Circulation Models (RRTMG) with aerosol source as prescribed monthly climatology from Advanced Very High-Resolution Radiometer (AVHRR) to develop the ensemble products⁹⁹. These products have been extensively utilized in previous studies for validation and comparative analyses, demonstrating consistent and stable quality across diverse time scales and under conditions of high meteorological variability^100,101,102. ERA-5 datasets are available at Copernicus Climate Data Store (https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels?tab=download). The temporal evolution of the planetary boundary layer height (PBLH) in Delhi during February and March, including comparisons with modeled data, is illustrated in Supplementary Fig. 11.

Calculation of the aerosol liquid water content

To calculate aerosol liquid water content (ALWC), we used the ISORROPIA II model. This model neglects the effects of curvature and surface tension and assumes that the uptake of aerosol water does not alter the ambient water vapor pressure, thus equating aerosol water activity (a_w) with relative humidity (RH). To account for different species, present in the particle phase, we used the Zdanovskii–Stokes–Robinson (ZSR) mixing rule, as shown in Ali et al.²⁴, yielding:

$$\text{ALWC}=\sum \frac{{\text{M}}_{\text{i}}}{{\text{m}}_{\text{oi}}\left({\text{a}}_{\text{w}}\right)}$$

(1)

where ALWC is estimated in kg/m³ air, M_i denotes the concentration of i^th component (mol/ m³ air), and m_oi (a_w) is the molality of an aqueous binary solution of the i^th electrolyte with the same a_w (i.e., relative humidity) as the multicomponent solution.

Weather research forecasting with coupled chemistry (WRF-Chem)

We employed the WRF-Chem v4.2 model^103,104 to investigate aerosol formation, distribution, meteorological parameters, and their dynamics over the study domain. Model simulations were performed using two nested domains with horizontal resolutions of 9 and 3 km, a temporal resolution of 1 h, and a vertical structure of 36 layers, including 22 layers within the first 1 km above ground level. The simulations incorporate optimized physical, dynamical, and radiative parameterizations (detailed in Table 1) following Srivastava et al.¹⁰⁵. Previous studies have demonstrated that WRF-Chem effectively captures the spatiotemporal variations of meteorological parameters, aerosols, and gas-phase chemistry over South and Southeast Asia, particularly during strong pollution events^106,107,108. Meteorological initial and boundary conditions were obtained from the National Centers for Environmental Prediction (NCEP) Global Data Assimilation System Final Analysis (GDAS/FNL) (National Centers for Environmental Prediction/National Weather Service/NOAA/U.S., 2015) at a spatial resolution of 0.25° × 0.25° and a temporal resolution of 3 h. Chemical boundary conditions were derived from the Community Atmosphere Model with Chemistry (CAM-Chem) on a 6 h temporal resolution. This study employs a coupled chemical mechanism for aerosol and gas-phase chemistry integrating the Goddard Chemistry Aerosol Radiation and Transport (GOCART) bulk aerosol scheme¹⁰⁹ and the Model for Ozone and Related Chemical Tracers (MOZART-4)^110,111. The scheme accounts for key aerosol species, including black carbon (BC), organic carbon (OC), sulfate, dust, and sea salt. Aerosol formation processes such as convection, advection, diffusion, emission, and dry deposition are incorporated. Additionally, this scheme comes with sulfur gas chemistry, where secondary aerosol species (sulfate) are formed through the oxidation of SO₂ and dimethyl sulfide (DMS) by hydroxyl radicals (OH), hydrogen peroxide (H₂O₂), and nitrate radicals (NO₃). Biogenic emissions of trace species were calculated online using the Model of Emission of Gases and Aerosols from Nature (MEGAN)¹¹². Biomass burning emissions were obtained from the Fire Inventory from NCAR (FINNv1.5)¹¹³, while anthropogenic emissions were sourced from the Emission Database for Global Atmospheric Research in collaboration with the Task Force for Hemispheric Transport of Air Pollution (EDGAR-HTAP)¹¹⁴.

Table 1 WRF-Chem model setup scheme