Diwali: Urban, Environmental & Cultural Insights

• Diwali is a major Indian festival characterized by intense firecracker displays, resulting in abrupt urban air quality and noise level changes.
• Studies use dense IoT sensor networks, chemical speciation, and spatial interpolation to capture sharp PM spikes and localized pollution during the festival.
• Modeling approaches, including epidemic simulations and cultural NLP datasets (DIWALI), leverage Diwali’s event-driven dynamics to analyze behavioral and cultural impacts.

Searching arXiv for the cited Diwali-related papers to ground the article in current records. Diwali is among the most important Indian festivals, and elaborate firework displays mark the evening’s festivities. In the recent technical literature, it appears as a recurrent, sharply episodic event whose signatures are observable in dense IoT sensing, chemical source apportionment, mobile noise monitoring, and agent-based epidemic simulation. The same token also appears as an acronym in NLP research—DIWALI, “Diversity and Inclusivity aWare cuLture specific Items for India”—which is a dataset for cultural text adaptation rather than a study of the festival itself (Manchanda et al., 2020, Parmar et al., 2022, Sahoo et al., 22 Sep 2025).

1. Diwali as a measurable urban event

The papers considered here describe Diwali through two operational features. First, they identify the festival as a period of heavy firecracker usage, especially in the evening. Second, they treat it as a predictable interval of altered human behavior, including festival intermingling, elevated ambient noise, and abrupt shifts in particulate matter. In Hyderabad, the event of interest was around the Diwali festival (4 Nov 2021), particularly the evening of heavy firecracker usage; in New Delhi, “during Diwali” is defined around the fireworks interval; and in Mumbai, the Diwali period is modeled as Nov 8 to Nov 14 with explicit changes to transmission parameters and compliance assumptions (Parmar et al., 2022, Manchanda et al., 2020, Harsha et al., 2020).

This framing matters methodologically. Diwali is not modeled as a slow seasonal trend but as a short-duration perturbation with high spatial heterogeneity and strong temporal synchronization. That is why the relevant studies use high-frequency acquisition, event-driven spatio-temporal analysis, chemical tracers, or explicit festival-period parameter changes rather than only coarse daily averages.

2. Dense PM monitoring and spatio-temporal signatures

A detailed IoT deployment in Hyderabad used 49 IoT-based PM monitoring devices, of which 43 were custom-built and 6 were commercial Airveda units, in a dense grid of approximately $4\,\mathrm{km}^2$. The devices were distributed across urban, semi-urban, green areas, and high-traffic junctions, and recorded PM2.5, PM10, temperature, and RH every 30 seconds continuously. Data from all nodes was aggregated on the ThingSpeak IoT cloud using secure MQTTS. Pre-deployment calibration co-located each low-cost PM sensor with an Aeroqual S500 for a week, yielding approximately 20,000 data points per device; simple linear regression was fit for each individual sensor and adjusted for seasonal effects,

$y(i) = m\,x(i) + c,$

and outliers were detected and removed using the Interquartile Range method,

$$L_r = Q_1 - 1.5\,I_{QR}, \quad U_r = Q_3 + 1.5\,I_{QR}.$$

Hourly-averaged time series and Inverse Distance Weighting spatial interpolation were then used to analyze Diwali evening conditions (Parmar et al., 2022).

The reported Diwali signature was a sudden rise after 8pm on 4 Nov 2021: a sharp and synchronized spike in PM10 and PM2.5 values across all locations immediately after 8pm, coinciding with the time of widespread firecracker usage. At device AV64, PM10 rose from 40 to 307 $\mu \mathrm{g/m}^3$, an increase by a factor of approximately 7.7. Elevated PM levels dropped rapidly in the following hours, which distinguishes the event spike from persistent elevation; rain earlier in the day and on following days reduced PM in the afternoon, but only the Diwali night showed the sharp, event-driven rise. Spatial interpolation at 17:00, 21:00, and 23:00 showed marked spread and peak immediately after firecracker activity, with localized regions of higher activity. The same study also showed that dense deployment resolved local PM hotspots, whereas sparse interpolations with 12 or 4 nodes underdetailed or entirely missed the event effect; the RMSE was 59.3 with 4 devices versus 32.5 with 12 devices (Parmar et al., 2022).

A plausible implication is that Diwali-related PM excursions are an identifiability problem as much as a sensing problem: when monitoring density is too low, the event can be statistically visible in aggregate while still being spatially mischaracterized.

3. Chemical speciation, tracers, and source apportionment of PM$_{2.5}$

In New Delhi, chemical speciation and source apportionment provide a more resolved view of Diwali aerosol composition. During fireworks, the average PM$_{2.5}$ concentration was approximately $192\,\mu\mathrm{g/m}^3$ at the IITMD site, representing a 408% increase over the pre-Diwali average, and the peak reached approximately $606\,\mu\mathrm{g/m}^3$, which is 16 times the pre-Diwali concentration. The elemental, organic, and black carbon fractions increased by factors of 46.1, 3.7, and 5.6, respectively; chloride increased by 7.4 times and sulfate-nitrate-ammonium by 2.5 times. The concentration of species such as K, Al, Sr, Ba, S, and Bi displayed distinct peaks during the firework event and were identified as tracers (Manchanda et al., 2020).

The source apportionment analysis used positive matrix factorization, expressed as

$X = GF + E.$

At each site, seven source categories were resolved: Fireworks, Biomass Burning, Coal Combustion, Industrial Emissions, Dust-related, Vehicular Emissions, and Secondary Chloride. During Diwali, fireworks accounted for 94% of elemental PM$_{2.5}$ at IITMD and 92% at IITD, compared with 3% and 6% pre-Diwali. The average fireworks-source concentration during Diwali was approximately 390 times versus ePD at IITMD and approximately 178 times at IITD. Following Diwali, fireworks still contributed approximately 54% of elemental PM$y(i) = m\,x(i) + c,$0 at IITMD and approximately 29.5% at IITD, but biomass burning became the principal driver for elevated PM$y(i) = m\,x(i) + c,$1 in the days after the festival (Manchanda et al., 2020).

This distinction addresses a common simplification. The studies do not treat “Diwali pollution” as a unitary phenomenon. Rather, they separate the acute firework episode from the subsequent haze regime. The paper states that Delhi has encountered serious haze events following Diwali in recent years, and that biomass burning emissions rather than the fireworks drive the poor air quality in the days following Diwali. Health implications are correspondingly species-specific: the average concentrations of potential carcinogens such as As exceeded US EPA screening levels for industrial air by a factor of approximately 9.6, while peak levels reached up to 16.1 times the screening levels (Manchanda et al., 2020).

4. Urban noise during Diwali

A separate Hyderabad study examined Diwali through mobile noise sensing. Measurement Campaign III was designed specifically to compare Diwali Day (Oct 31, 2024) with a Typical Day (Nov 6, 2024) on a route covering IIIT Hyderabad to interior residential areas including SR Nagar and Sanath Nagar. Sampling lasted approximately 4.5 hours per day over roughly 70 km per day, and more than 29,928 data points were collected at 1-second intervals and later averaged to 10 seconds for analysis. The study reported that a clear spike in noise levels begins at 7:00 PM, peaks between 7–9 PM, and remains elevated compared to the typical day during those hours (Manthina et al., 31 Aug 2025).

The quantitative contrast is explicit. The mean noise level on Diwali was 80.2 dBA and the variance was 88.9, whereas the Typical Day had a mean of 74.1 dBA and variance of 57.1; the maximum observed on Diwali was 110 dBA. Spatial maps showed intense hotspots in specific areas, mostly residential, where celebrations are more significant, and these patterns were less pronounced or absent on the typical day. For calibration during mobile measurements, Random Forest Regression achieved $y(i) = m\,x(i) + c,$2 and RMSE $y(i) = m\,x(i) + c,$3 without velocity, improving to $y(i) = m\,x(i) + c,$4, RMSE $y(i) = m\,x(i) + c,$5, and MAE $y(i) = m\,x(i) + c,$6 when vehicular velocity was included (Manthina et al., 31 Aug 2025).

The paper also places these measurements against regulatory thresholds: CPCB standards for residential areas are 55 dBA by day and 45 dBA by night, while silence zones are 50 dBA by day and 40 dBA by night. During Diwali, mean and peak noise levels far exceeded these legal standards. This suggests that Diwali is not only an aerosol event but also an acoustically extreme interval whose spatial structure is similarly heterogeneous and hotspot-driven.

5. Diwali in epidemic simulation

Diwali also appears in epidemic modeling as a period of increased intermingling. The IISc-TIFR agent-based simulator for Mumbai constructs a synthetic population of 12.8 million and models household, workplace, school, community, and transit interactions. To capture Ganpati, Navratri/Dussehra, and Diwali, the simulations increase the community transmission rate by two-thirds and reduce compliance with distancing and masking from 60% to 40% in non-slums and from 40% to 20% in slums during the festival weeks. For Diwali specifically, the period is Nov 8 to Nov 14 (Harsha et al., 2020).

Under these assumptions, the model indicates that the effect of Diwali is contingent rather than unconditional. If the intermingling level during Navratri/Dussehra and Diwali festival is similar to Ganpati, then, because these festivals occur later when a larger fraction of the population has already been infected, the resulting infections are likely to be less and the overall impact on the city’s medical infrastructure relatively less. The projected prevalence by mid-January 2021 stabilizes at approximately 80% in slums and approximately 55% in non-slums; fatalities stabilize by March 2021 around 13,000–14,000, and cumulative infections stabilize between 8–9 million residents. The time series show noticeable but manageable surges during each festival relaxation including Diwali. At the same time, the paper states that if there is a substantial increase in interaction and the social distancing and mask-related precautions are relatively weakened, then one may again see a significant rise in infections (Harsha et al., 2020).

This is a model-based result, not a direct measurement study. Its significance lies in formalizing Diwali as a structured change in contact behavior. The same festival interval that is observed through PM and noise sensors is here encoded through increased $y(i) = m\,x(i) + c,$7 and reduced compliance, linking cultural calendars to epidemic dynamics.

6. DIWALI as an NLP acronym for cultural text adaptation

In computational linguistics, DIWALI denotes a distinct construct: “Diversity and Inclusivity aWare cuLture specific Items for India.” This dataset contains 8,817 cultural concepts spanning 17 cultural facets and 36 sub-regions, defined as 28 states plus 8 union territories. The facets include Clothing, Textiles, Jewellery, Food, Drinks, Festivals, Rituals, Traditions, Dance Forms, Traditional Games, Religion, Arts, States & Capitals, Places, Languages & Dialects, Architectures, and Names. The curation pipeline began with GPT-4o prompting and Wikipedia search, which was judged insufficient for coverage, and the final curation combined LLM output with authoritative web searches and manual verification for link accuracy and cultural validity (Sahoo et al., 22 Sep 2025).

The dataset is used for a cultural text adaptation task: adapting text from American to Indian context without loss of semantic intent but maximizing cultural resonance. Evaluation combines a CSI Adaptation Score, LLM as Judge, and human evaluation. The automatic score is defined as

$y(i) = m\,x(i) + c,$8

Seven open-weight LLMs from three families were evaluated across GSM8k, MGSM, DailyDialog, and ROCStories, using both English and Bengali prompts. DIWALI is explicitly compared with CANDLE and DOSA; it contains 8,817 concepts, versus approximately 650 Indian concepts for CANDLE and approximately 615 items across 19 states for DOSA. For GSM8k with English prompts, DIWALI-based exact-match AAS values reported in the paper include 0.855 for Llama-2-7B-chat-hf, 0.605 for Llama-3.1-8B-Instruct, 0.933 for Llama-3.2-1B-Instruct, 0.672 for Llama-3.2-3B-Instruct, and 0.563 for Mistral-7B-Instruct (Sahoo et al., 22 Sep 2025).

The paper’s substantive conclusion is not that current models achieve deep cultural competence. Human evaluators from Chhattisgarh, Maharashtra, Kerala, Bengal, and Delhi rated most models as only moderate or surface-level in cultural adaptation, and no model achieved consistent deep integration. LLM judges assigned systematically higher Cultural Relevance scores than humans, with inflation by +0.5 to +2.5 points. Heatmap analyses further showed selective sub-regional coverage, with heavy clustering in populous or better-known states and near-total neglect of the Northeast or smaller regions. Here “DIWALI” is a namesake acronym rather than the festival, but its inclusion is notable because it treats Indian cultural specificity—including festivals—as a measurable evaluation target (Sahoo et al., 22 Sep 2025).