AirIndex: Multi-Domain Index Constructs

• AirIndex is a family of context-dependent index constructs that span environmental air quality indicators and storage-optimized hierarchical indexes.
• Environmental applications of AirIndex utilize diverse sensing modalities and calibration methods for low-cost monitoring and fine-grained forecasting.
• In database research, AirIndex refers to an I/O-aware index builder that significantly reduces lookup latency compared to traditional methods.

AirIndex is a label used in the literature for multiple, non-equivalent constructs rather than a single standard. In the cited works it denotes a simplified MQ-135-based air-quality indicator, CPCB- and EPA-aligned AQI visualizations and predictors, the Air Stagnation Index used for seasonal meteorological forecasting, fine-grained urban AQI mapping and forecasting services, composite and uncertainty-aware indices for outdoor or indoor air assessment, and, in database systems, an I/O-aware builder for hierarchical lookup indexes (Karar et al., 2020, Zhou et al., 2023, Sharma, 12 Jun 2025, Ha et al., 2020, Chockchowwat et al., 2023).

1. Terminological scope and naming variants

The term is therefore best understood as a family of context-dependent index constructs. In atmospheric and environmental papers, it usually refers either to an air-quality indicator itself or to a framework that estimates, visualizes, or forecasts such an indicator. In one systems line of work, however, AirIndex denotes a storage-optimized hierarchical data structure builder rather than an environmental metric (Chockchowwat et al., 2023).

Usage domain	Meaning of “AirIndex”	Representative work
Low-cost sensing	Single-sensor categorical indicator from MQ-135 PPM	(Karar et al., 2020)
Regulatory visualization	CPCB AQI and pollutant sub-index display in AR	(Mathews et al., 2020)
Meteorology	Air Stagnation Index for pollutant dilution conditions	(Zhou et al., 2023)
Fine-grained forecasting	Neighborhood-scale AQI or PM forecasting service	(Sharma, 12 Jun 2025)
Composite/uncertainty-aware indexing	Similarity, fused indoor index, or fuzzy AQI	(Saxena et al., 2019, Ha et al., 2020, Inzmam et al., 20 Mar 2026)
Database systems	I/O-aware hierarchical lookup index builder	(Chockchowwat et al., 2023)

A further naming complication is explicit in the database literature: the 2022 paper describes the system as AutoIndex, while the 2023 paper uses AirIndex for the same general problem class of tuning hierarchical indexes through data and storage characteristics (Chockchowwat et al., 2022). This terminological overlap is important because it separates environmental “index” semantics from storage-engine “index” semantics.

A recurrent misconception in the air-quality papers is to assume that any construct called AQI or AirIndex implements a regulatory multi-pollutant AQI exactly. That is not uniformly true. Some works port CPCB or EPA-style sub-index logic; others use a single pollutant, a proxy variable, or a custom aggregation rule (Karar et al., 2020, Mathews et al., 2020, Rowley et al., 2022).

2. AirIndex as an air-quality indicator in monitoring systems

In low-cost embedded monitoring, AirIndex can be a direct categorical mapping from a sensor output rather than a standards-compliant AQI. GASDUINO defines the Air Quality Index as a single-sensor, single-number indicator expressed in PPM from the MQ-135 and mapped by IF–THEN rules into three categories: Good for 0 to 50 PPM, Moderate for 51 to 150 PPM, and Unhealthy for 151 to 200 PPM and above, with green, blue, and red indicators respectively. The paper explicitly states that it does not implement the EPA-style AQI computation, does not provide breakpoint interpolation, and does not disclose a calibration model from raw MQ-135 readings to pollutant-specific PPM; the index is thus a direct categorical proxy derived from a broad-sensitivity gas sensor (Karar et al., 2020).

By contrast, AiR treats AirIndex as India’s National Air Quality Index rendered into an augmented-reality interface. The application uses CPCB measurements, detects the nearest monitoring station from GPS with an optimized Haversine approach, computes pollutant sub-index values using CPCB breakpoints, and displays a cumulative AQI score together with pollutant-specific information for PM10, PM2.5, NO2, SO2, CO, O3, and NH3. The paper emphasizes that the computation was implemented in C# by “porting the system used by CPCB,” and that the result is visualized through a radial gauge and pollutant-specific panels rather than through a new AQI definition (Mathews et al., 2020).

Other device-centric implementations sit between those two extremes. The portable Bluetooth system of 2025 measures PM2.5, PM10, and CO, computes sub-indices with the standard piecewise linear AQI equation, and takes the maximum sub-index as the final AQI, but it does not explicitly identify the breakpoint standard used in code and appears to compute AQI from instantaneous readings rather than from NowCast or regulatory averaging windows. The smartphone-image work for Dhaka trains a DCNN to predict PM2.5 from outdoor images and then derives an equivalent daily averaged AQI aligned with Bangladesh category bands from Good to Extremely Unhealthy; here the reported “AirIndex” is image-mediated and PM2.5-centered rather than station-native (Rahman et al., 17 Jun 2025, Mondal et al., 2023).

Taken together, these implementations show that device-level AirIndex systems span at least three regimes: direct proxy mapping, standards-porting, and inferred AQI from an intermediate predicted pollutant. This suggests that the operational meaning of the index is determined as much by sensing and calibration choices as by the nominal category labels.

3. Forecasting and mapping formulations

In forecasting research, AirIndex often denotes a target field or trajectory rather than a directly measured scalar at a single site. The clearest meteorological example is the Air Stagnation Index in southern China. That study constructs a daily binary stagnation indicator from ERA5 variables including 10 m and aloft wind speeds, daily accumulated precipitation, planetary boundary layer height, CAPE, and CIN, then aggregates it to monthly and winter seasonal ASI. Autumn Niño indices are negatively correlated with wintertime ASI in the seven-province southern China region, with reported correlations of $-0.42$, $-0.50$, $-0.46$, and $-0.53$ for Niño 1+2, 3, 3.4, and 4 respectively. An LSTM using past ASI plus autumn Niño indices achieves the best reported validation performance at $k = 9$, with Corr $= 0.778$ and MAPE $= 0.143$ over 2000–2020 (Zhou et al., 2023).

At finer urban scales, the term is used for neighborhood-level forecasting services. The AirDelhi study frames the problem on 1 km$^2$ neighborhoods and 30-minute bins, using mobile sensors mounted on public buses, inverse-distance weighting imputation with $K=3$ and $p=3$, and spatio-temporal GNNs or GRUs to forecast PM2.5 and PM10 that can be mapped to AQI categories. Reported 24-hour test performance includes RMSE $-0.50$0, $-0.50$1, and AQI-category accuracy $-0.50$2 for PM2.5 with “GCN Image,” while GRU shows notable robustness on unseen coordinates in the Extended setting (Sharma, 12 Jun 2025).

A related urban mapping formulation appears in the Lahore case study, where 30 AQI sensors supervise estimation over 2,401 grid cells at 1 km resolution and hourly cadence. The framework fuses MODIS MAIAC AOD, MODIS LST, ERA5 meteorology, OpenStreetMap roads and green spaces, Google Maps PoIs, and LandScan population in a CNN+GCN architecture with a graph smoothness regularizer. The best average result over 10 runs is reported at learning rate $-0.50$3 and $-0.50$4, with MAE $-0.50$5, RMSE $-0.50$6, and MAPE $-0.50$7 (Ahmad et al., 20 Jan 2025).

More explicitly causal or physics-guided forecasting variants extend this logic. AirCade treats AQI trajectories and meteorological covariates as distinct causal objects, decouples synchronous and lagged effects, uses DK-Prompt embeddings and intervention masks for uncertain future meteorology, and reports over $-0.50$8 relative improvement on MAPE against the best baseline on KnowAir. The Dallas County benchmark instead adds EPA breakpoint-based AQI consistency as a weighted loss term in MLP+Physics and LSTM+Physics models for PM2.5 and O3 over horizons of 1, 7, 14, and 30 days, with the clearest gains for PM2.5 and shorter horizons (Ma et al., 26 May 2025, Masud et al., 22 Mar 2026).

4. Composite, similarity-based, and uncertainty-aware index constructions

Not all AirIndex formulations follow the conventional “maximum sub-index” paradigm. Several papers instead define explicitly composite indices.

Index	Aggregation rule	Distinguishing feature
DSI	Geometric mean of O3, SO2, CO similarities	Similarity to Delhi rather than direct health AQI (Saxena et al., 2019)
EIAQI	$-0.50$9	Fuses indoor pollutant exposure and humidex (Ha et al., 2020)
$-0.46$0 in AQNet	$-0.46$1	Custom WHO-threshold-based composite, not EPA/EU AQI (Rowley et al., 2022)
Weighted IT2 fuzzy IND-AQI	Type-2 fuzzy inference plus IT2-FAHP weights	Uncertainty-aware CPCB AQI enhancement (Inzmam et al., 20 Mar 2026)

The Delhi Similarity Index is the most explicitly comparative formulation. It transforms annual mean O3, SO2, and CO concentrations into pollutant-wise similarity scores relative to Delhi using a Bray–Curtis-like term raised to pollutant-specific exponents, then aggregates them by geometric mean. On that scale, Bengaluru is reported at roughly $-0.46$2 to $-0.46$3 over 2011–2014, while Jungfraujoch is reported around $-0.46$4 to $-0.46$5. The paper is explicit that NO2 could not be incorporated because the weight exponent computation did not yield a valid result under the dataset conditions used (Saxena et al., 2019).

Indoor air-quality work generalizes the notion further by fusing pollutant and comfort dimensions. The EIAQI system computes pollutant sub-indices for CO2, CO, H2, NH3, ethanol, H2S, toluene, and O2, computes humidex from temperature and relative humidity, and combines them through $-0.46$6. The fusion is stabilized by an Extended Fractional-Order Kalman Filter with a Matérn covariance prior, implemented with Waspmote sensors and the FOMCON toolbox (Ha et al., 2020).

The ontology-based 2026 framework stays within CPCB’s IND-AQI scale but replaces crisp breakpoints with interval Type-2 fuzzy sets and derives pollutant importance weights by IT2-FAHP. The final pollutant weights are reported as PM2.5 $-0.46$7, PM10 $-0.46$8, CO $-0.46$9, O3 $-0.53$0, NO2 $-0.53$1, SO2 $-0.53$2, and NH3 $-0.53$3, with AQI classes Good, Satisfactory, Moderate, Poor, Very Poor, and Severe preserved at the output level. The paper reports accuracy near $-0.53$4, precision near $-0.53$5, F1-score near $-0.53$6, and AQI category-prediction MAE $-0.53$7, RMSE $-0.53$8 relative to CPCB ground truth (Inzmam et al., 20 Mar 2026).

A comparable departure from regulatory AQI appears in AQNet, where predicted concentrations of NO2, O3, and PM10 are aggregated into a unitless index $-0.53$9 using WHO-informed thresholds of $k = 9$0, $k = 9$1, and $k = 9$2. The paper states explicitly that this is not US EPA or EU AQI scaling; $k = 9$3 indicates unhealthy levels on average, while $k = 9$4 indicates pollution present but not necessarily threshold exceedance (Rowley et al., 2022).

5. Sensing modalities, data fusion, and operational pipelines

The diversity of AirIndex definitions is mirrored by equally diverse sensing infrastructures. GASDUINO exemplifies a minimal IoT pipeline: Arduino UNO reads MQ-135, ESP8266 transmits data to the Remote XY cloud, and an Android UI displays the current PPM and category color. The processing path is simply sensor readout, unspecified PPM interpretation, IF–THEN classification, and transmission to the user interface (Karar et al., 2020).

AiR uses a very different architecture. It is built in Unity with AR Foundation, collects CPCB data through a Python-based pipeline into SQL storage, serves processed station histories through a REST API as JSON, and renders pollutant-specific 3D objects in a $k = 9$5 airspace around the user. Missing data longer than 2–3 days were removed after rechecking, shorter gaps were linearly interpolated, and the application distinguishes live nearest-station views from user-selected simulated localities (Mathews et al., 2020).

UAV-centric systems push the index into meter-scale spatial profiling. ARMS mounts a Plantower laser-based detector on a DJI Phantom 3, samples a 5 m grid with 10-second hovering at each waypoint, and reconstructs fine-grained AQI maps with the hybrid GPM-NN model and an adaptive, battery-aware monitoring algorithm. Over 100 days of measurements in a 2D roadside park and a 3D courtyard, the system reported higher prediction accuracy than linear interpolation and multi-variable linear regression while substantially reducing power consumption at suitable PDT settings (Yang et al., 2017).

ImgSensingNet couples UAV vision with a 3D wireless sensor network rather than using the UAV as the pollutant sensor. It collected 17,630 photos and 2.6 million AQI samples, inferred regional AQI scales from haze images using a 3D CNN over six haze-relevant feature maps, and used those priors to decide which ground sensors to wake via a joint estimation error criterion and a minimum independent dominating set heuristic. The paper reports overall energy-consumption reduction of $k = 9$6 with inference accuracy $k = 9$7 (Yang et al., 2019).

Indoor deployments add yet another sensing topology. The enhanced indoor system uses Waspmote nodes and Meshlium gateways, senses “round the clock,” and fuses gas concentrations and humidex through an EFKF. The case study identifies 16 sensors on one floor and more than 100 across the building, underscoring that AirIndex may be a building-management signal as much as a public outdoor one (Ha et al., 2020).

Across these architectures, AirIndex is inseparable from acquisition design: waypoint spacing, temporal resolution, wake-up policy, imputation, calibration, and communication protocol all materially affect what the index means and how stable it is operationally.

6. AirIndex in database systems

In database and storage systems, AirIndex designates a fundamentally different object: a data-and-I/O-aware index builder for hierarchical lookup structures. The 2023 paper formalizes a class of layered indexes in which each layer maps a search key to a byte-range in the next layer, models storage transfer time as $k = 9$8, typically with the affine profile $k = 9$9, and minimizes expected end-to-end lookup latency over the number of layers, layer types, and per-layer materialization (Chockchowwat et al., 2023).

The system supports exact step nodes and approximate band nodes, builds candidate layers with GStep, GBand, and EBand builders, and searches the exponentially large design space with the graph-based optimizer AirTune. The lookup objective is expressed as the read time of the full root layer plus the read times of partial ranges fetched in each lower layer. In the reported experiments, AirIndex delivers up to $= 0.778$0 faster lookup than LMDB, $= 0.778$1–$= 0.778$2 faster lookup than RMI/CDFShop, PGM-Index, ALEX/APEX, and PLEX, and $= 0.778$3 faster lookup than Data Calculator’s suggestion across datasets and storage settings (Chockchowwat et al., 2023).

The 2022 related paper presents the same basic idea under the name AutoIndex and emphasizes joint optimization of structure and internal regressors for local SSD and Azure Cloud Storage. There the reported gains are $= 0.778$4–$= 0.778$5 faster lookup on local SSD and $= 0.778$6–$= 0.778$7 faster lookup on Azure Cloud Storage relative to state-of-the-art methods, with structure adapting from tall-and-narrow at short RTT to shallow-and-wide at large RTT (Chockchowwat et al., 2022).

The coexistence of this systems meaning with the environmental meanings is not merely terminological trivia. It establishes that “AirIndex” has no discipline-independent definition. In atmospheric science, public-health informatics, IoT sensing, and GIS, it usually denotes an air-quality or air-stagnation indicator. In database research, it denotes an optimized hierarchical lookup index. Any technical reading of the term therefore requires immediate attention to domain, aggregation rule, sensing pipeline, and standards alignment.