AQUAIR: Indoor IEQ Data for Smart Aquaculture
- AQUAIR is an open, high-resolution indoor IEQ dataset that records temperature, humidity, CO₂, VOC, PM2.5, and PM10 in a trout hatchery room.
- The dataset supports short-horizon forecasting, event detection, and anomaly analysis, making it a benchmark for smart aquaculture research.
- A structured preprocessing pipeline and rigorous quality control ensure AQUAIR’s reliability as a resource for linking head-space conditions to aquaculture performance.
Searching arXiv for the AQUAIR dataset paper and adjacent AQUAIR-related usages to ground the article in current papers. AQUAIR is an open, high-resolution indoor environmental quality (IEQ) dataset collected above trout tanks in an indoor recirculating aquaculture facility and intended for smart aquaculture research, especially forecasting, event and anomaly detection, and studies of air–water interactions in fish farming environments. The dataset logs six IEQ variables—air temperature, relative humidity, carbon dioxide, total volatile organic compounds, PM2.5, and PM10—at 5-minute resolution over 84 days in a trout hatchery room at Amghass Station 3, Azrou, Morocco, and is publicly archived on Figshare under CC-BY-4.0 (Sabiri et al., 28 Sep 2025).
1. Definition and scientific scope
AQUAIR is a public dataset of indoor air conditions measured inside a trout hatchery room. It records air temperature, relative humidity, CO₂, total VOCs, PM₂.₅, and PM₁₀ at 5-minute resolution over 84 days, and its stated purposes are to provide a realistic, continuous IEQ time series for smart aquaculture monitoring, a benchmark dataset for time-series ML, and a resource for linking head-space air conditions to water quality and fish welfare in recirculating aquaculture systems (Sabiri et al., 28 Sep 2025).
The dataset addresses a specific gap in aquaculture informatics. Most aquaculture datasets focus only on water quality, whereas AQUAIR provides detailed head-space air data in an actual trout rearing room. In the context described for indoor RAS, head-space monitoring matters because gas exchange links air CO₂ to dissolved CO₂ and pH, humidity and temperature affect evaporation, dissolved oxygen solubility, and biofilter performance, and VOCs and particulates may reflect disinfectants, feed, materials, feeding, cleaning, and human activity. This framing makes AQUAIR a head-space dataset rather than a general aquarium or water-quality dataset.
A recurrent misconception is to treat AQUAIR as a direct fish-welfare or water-chemistry dataset. The paper is explicit that AQUAIR itself contains only air data. Its contribution is to make head-space dynamics measurable at a cadence and duration suitable for coupling with water-quality dynamics in subsequent work, not to provide those water variables directly (Sabiri et al., 28 Sep 2025).
2. Measurement environment and instrumentation
The measurements were acquired in a closed indoor trout room in a public inland-fish facility operated by the Centre National d’Hydrobiologie et de Pisciculture. The site is Amghass Station 3 near Amghass, Azrou, Morocco, at coordinates 33°23′37.0″ N, 5°27′01.9″ W and elevation ~1250 m, in a cool-summer Mediterranean climate with mean annual temperature ≈ 12 °C and ≈ 650 mm precipitation. The room is ~75 m², contains five large trout ponds of ≈ 4 m³ each supplied by a common RAS loop, has no windows, and is mechanically ventilated by a small extractor fan above the ponds. Air temperature was actively controlled between ~6 and 22 °C during the campaign, and occupancy increased during feeding, cleaning, and health checks (Sabiri et al., 28 Sep 2025).
A single Awair HOME monitor (AQM-8002A) was used. The integrated sensors cover temperature, relative humidity, CO₂, VOC, and PM₂.₅/PM₁₀. The stated ranges and accuracies are: temperature, −40 to 125 °C with ±0.3 °C; relative humidity, 0–100% with ±2% RH; CO₂, 400–5000 ppm with ±75 ppm or 10%; VOC, 20–36 000 ppb with ±15%; and PM₂.₅ / PM₁₀, 0–1000 µg m⁻³ with ±15 µg m⁻³ or 15%. The sensing principles summarized in the paper are an integrated digital sensor for temperature and RH, NDIR for CO₂, a metal-oxide or equivalent tVOC element for VOC, and laser scattering with a small fan for PM (Sabiri et al., 28 Sep 2025).
Sensor placement followed ISO 16000-1 recommendations. The monitor was mounted 1.5 m above the floor, horizontally centered between the two rows of tanks, at least >1 m from the nearest water surface, and in an unobstructed central aisle to ensure sufficient airflow (≥ 0.1 m s⁻¹) for the PM sensor. Power was supplied via a surge-protected 5 V brick with cable routing away from splashing. This arrangement was intended to represent room-typical head-space conditions rather than boundary layers directly over water or poorly ventilated corners.
Data acquisition used the Awair cloud / Home App rather than the local API, with cadence configured to 5 minutes and export performed as CSV through the cloud dashboard. The paper also notes calibration checks against reference instruments, emphasizes internal consistency and agreement with expected ranges for a controlled hatchery, and validates the dataset using descriptive statistics, extremes analysis, diurnal cycles, Spearman correlation structure, and completeness visualization (Sabiri et al., 28 Sep 2025).
3. Dataset structure, packaging, and archival form
The dataset consists of timestamped IEQ snapshots stored in UTC at 5-minute cadence. The full column structure is shown below.
| Column | Meaning | Unit / format |
|---|---|---|
timestamp(UTC) |
ISO 8601 timestamp | UTC |
score |
Awair proprietary IEQ index | 0–100, dimensionless |
temp |
Air temperature | °C |
humid |
Relative humidity | % |
co2 |
CO₂ concentration | ppm |
voc |
Total volatile organic compounds | ppb |
pm25 |
PM₂.₅ | µg m⁻³ |
pm10 |
PM₁₀ | µg m⁻³ |
The time span runs from 14 October 2024 to 9 January 2025. The cadence is 5 minutes, corresponding to 12 samples per hour, and the total coverage is 84 days with a known maintenance gap. There is no data from 11–14 December 2024 due to scheduled sensor maintenance; the paper characterizes this gap as structural rather than environmental (Sabiri et al., 28 Sep 2025).
After quality control, the dataset contains 23 856 5-minute records. It is distributed as two UTF-8 CSV files, AQUAIR_1.csv covering 14 Oct 2024 to 10 Dec 2024 and AQUAIR_2.csv covering 15 Dec 2024 to 09 Jan 2025. Both files contain the full set of columns. Units are SI-consistent; particulates are stored as whole-number µg m⁻³ and gases with up to four decimals.
The archive is hosted on Figshare under CC-BY-4.0 at DOI 10.6084/m9.figshare.28934375.v1. The license permits reuse, including commercial reuse, with attribution. The paper also describes the result as fully quality-controlled and analysis-ready, which is central to its role as a benchmark dataset rather than a raw device dump (Sabiri et al., 28 Sep 2025).
4. Processing pipeline and quality control
AQUAIR is accompanied by a structured preprocessing workflow intended to generate analysis-ready time series. The first stage is timestamp normalization. Timestamps are converted to ISO 8601, and although the text notes a minor inconsistency in formatting examples, the stored timestamps are explicitly UTC and use the Z indicator in the data records section (Sabiri et al., 28 Sep 2025).
The second stage is 5-minute grid anchoring. Data are aligned to an exact 5-minute grid at minutes 00, 05, 10, …, 55. If multiple readings fall into the same 5-minute slot, they are averaged. Missing slots are explicitly represented as gaps, with no synthetic filling at that stage. This preserves structural outages, including the maintenance interval, as part of the temporal record rather than obscuring them.
Missing-value treatment is variable-specific. Entire missing rows are left as gaps. Single-channel gaps of ≤10 minutes, defined as up to two consecutive missing samples in one variable while other channels exist at that timestamp, are filled by linear interpolation. Longer gaps remain NaN and are flagged. The interpolation rule is given as
for , where and are the nearest valid values bracketing the missing time (Sabiri et al., 28 Sep 2025).
Outlier detection uses a rolling Hampel filter with window and threshold at . A point that deviates from the local median by more than 3 local standard deviations is treated as an outlier, replaced by the local median, and flagged in a qa_flag bitmask that is referenced conceptually but not exposed as a separate column in the provided description. After outlier correction, range tests enforce the manufacturer’s operating limits, including PM ≤ 1000 µg m⁻³, CO₂ within 400–5000 ppm, VOC within 20–36 000 ppb, and temperature and RH within their stated ranges. Violations are set to NaN and flagged (Sabiri et al., 28 Sep 2025).
The final stage is unit harmonization: temperature in °C, RH in %, CO₂ in ppm, VOC in ppb, and PM in µg m⁻³ as integers. The paper mentions an open-source processing pipeline, although the GitHub URL is not explicitly provided in the text block. A plausible implication is that reproducibility is intended to rest on the explicitly described steps together with the archived files, rather than on an opaque proprietary export process.
5. Statistical profile and temporal organization
The descriptive statistics reported after QC show a relatively constrained thermal regime with stronger episodic variation in gas and particulate channels. For temperature, the paper reports mean 15.95 °C, SD 1.75, median 16.2, and min/max 11.7 / 20.7. Relative humidity has mean 75.79%, SD 17.50, median 86.9, and min/max 48.2 / 93.3. CO₂ has mean 1143 ppm, SD 804, median 758, and min/max 400 / 3704. VOC has mean 469 ppb, SD 640, median 143, and min/max 20 / 9186. PM₂.₅ has mean 18.28 µg m⁻³, SD 20.97, median 12.2, and min/max 0 / 505, while PM₁₀ has mean 19.58 µg m⁻³, SD 21.38, median 13.2, and min/max 1 / 513 (Sabiri et al., 28 Sep 2025).
These values support the paper’s characterization of generally stable conditions with event-driven excursions. Temperature is tightly clustered around ~16 °C, and humidity is frequently above 80%, consistent with an indoor room containing open water surfaces. By contrast, CO₂ and VOC have wide upper percentiles, and particulates exhibit occasional spikes into the ~500 µg m⁻³ range. The paper’s boxplot-based discussion states that no implausible extremes remain after QC and that CO₂ and VOC exhibit the widest dynamic ranges, dominated by operations.
The mean diurnal cycle is operationally structured. Temperature shows a clear mid-afternoon peak around 16.4 °C, while relative humidity dips when temperature rises. CO₂, VOC, PM₂.₅, and PM₁₀ show strong morning peaks from 08:00–10:00 tied to feeding and evening peaks from 18:00–20:00 associated with cleaning and maintenance. This pattern is one of the main reasons the dataset is positioned as suitable for short-horizon forecasting and event detection rather than only static environmental description (Sabiri et al., 28 Sep 2025).
The correlation structure reinforces that interpretation. The paper reports Spearman for temperature versus RH, for PM₂.₅ versus PM₁₀, and for CO₂ versus VOC. The first is consistent with inverse temperature–humidity behavior in relative terms, the second with a shared particle-generation process, and the third with common operational drivers such as feeding, cleaning, and occupancy.
The completeness discussion requires careful reading. Table 3 reports 5.26% missing per channel, while Figure 1 separately reports overall completeness ≈ 99.7% and visualizes the principal 11–14 December maintenance gap together with a few brief sub-hour outages. The paper does not resolve this discrepancy in the excerpted text, so both values are best treated as reported descriptors attached to different validation views rather than as a single harmonized completeness metric (Sabiri et al., 28 Sep 2025).
6. Research uses, limitations, and terminological ambiguity
AQUAIR is explicitly positioned as a benchmark for short-horizon forecasting, event detection, anomaly detection, environmental sensing research, and data-centric machine learning education. The stated model families include classical time-series methods such as ARIMA, SARIMA, and Prophet, and ML/DL methods such as random forests, LSTM, GRU, transformers, and hybrid models. The dataset supports both univariate and multivariate forecasting, and the paper notes 30–60 minute-ahead prediction of variables such as CO₂, PM₂.₅, and temperature as natural task formulations. It is also proposed for detecting feeding events, cleaning and maintenance windows, ventilation or occupancy changes, and point, contextual, or collective anomalies (Sabiri et al., 28 Sep 2025).
The paper also frames AQUAIR as a resource for coupling with external data sources. Suggested mergers include water quality time series such as temperature, DO, pH, ammonia, and nitrogen; fish health records such as mortality, disease outbreaks, and growth; and behavioral data such as video-based activity and feeding responses. This suggests a broader role for AQUAIR as a head-space component within multimodal RAS monitoring pipelines rather than as a complete aquaculture state description.
Its limitations are explicitly stated. AQUAIR comes from a single site, one room, and one Awair device, which limits generalizability across systems, climates, and hardware. The time span is 84 days from mid-October to early January, so it does not cover a full annual cycle or different farming stages. It contains no concurrent water-quality or fish-behavior data. The hardware is a consumer-grade IEQ device with potential biases and a single-point spatial sample, and the record includes a scheduled maintenance gap and small sub-hour outages. These caveats matter methodologically: the dataset is well suited to benchmarking temporal models on realistic IEQ data, but not to claims about universal indoor aquaculture conditions (Sabiri et al., 28 Sep 2025).
A further source of confusion is terminological. In adjacent robotics literature, “AQUAIR” typically refers to hybrid aerial–aquatic vehicles with dual-medium propulsion, cross-interface dynamics, and sensing and control across air and water, as stated in the simulator paper "AAM-SEALS: Developing Aerial-Aquatic Manipulators in SEa, Air, and Land Simulator" (Yang et al., 2024). Related examples include a variable-stiffness aerial–aquatic locomotion robot with a biomimetic propulsion module (Hu et al., 2024) and the quadrotor platform "TJ-FlyingFish: Design and Implementation of an Aerial-Aquatic Quadrotor with Tiltable Propulsion Units" (Liu et al., 2023). In a different adjacent usage, an IoT smart aquarium paper presents an “AQUAIR-type system” for real-time water-quality monitoring and automated feeding (Ayon et al., 13 Jan 2026). Within (Sabiri et al., 28 Sep 2025), however, AQUAIR denotes a high-resolution IEQ dataset for indoor aquaculture head-space monitoring, not an aerial–aquatic robot and not a smart aquarium controller.
Taken in that narrower sense, AQUAIR fills a specific infrastructural role in smart aquaculture: it provides a reproducible, openly licensed, quality-controlled record of head-space dynamics above trout tanks in an indoor RAS environment, with enough temporal structure to support forecasting, event characterization, anomaly analysis, and methodological work on environmental sensing (Sabiri et al., 28 Sep 2025).