📄 earth_science

An air quality, meteorological and traffic spatio-temporal annotated data set

This paper presents a six-year, hourly spatio-temporal dataset from Madrid that integrates air quality, meteorological, and traffic data from over 5,000 sensors to support urban environmental and mobility research.

Original authors: David Maria-Arribas, Alejandra Abalo-García, Antonio S. Montemayor, Juan J. Pantrigo, Alfredo Cuesta-Infante

Published 2026-07-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: David Maria-Arribas, Alejandra Abalo-García, Antonio S. Montemayor, Juan J. Pantrigo, Alfredo Cuesta-Infante

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the air around us as a giant, invisible soup. Sometimes this soup is clear and fresh, but other times, it gets thick with invisible ingredients like car exhaust, dust from tires, and smoke from factories. This is what scientists call "air pollution." But this soup doesn't just sit there; it changes constantly. The wind might blow the smog away, or a heavy rain might wash it out. The sun might heat it up, making it react in new ways. This is "meteorology," or the study of weather. And what stirs the soup the most? Us. Specifically, our cars. When traffic is heavy, the soup gets thicker with pollution; when the roads are empty, it clears up. This is the connection between "traffic" and "air quality."

For a long time, scientists have tried to understand this relationship, but they often had to look at the ingredients separately. They might have a recipe for the weather, a list of the pollution, and a log of the traffic, but these lists were often written in different languages, at different speeds, and for different places. It's like trying to bake a cake when your flour is measured in cups, your sugar in grams, and your eggs in dozens, all written on different scraps of paper from different years. To really understand how our cities breathe, we need to mix these ingredients together into one perfect, synchronized recipe. That is exactly what this new study has done.

The paper introduces a massive new digital collection called METRAQ, which acts like a giant, time-traveling library for the city of Madrid, Spain. Think of it as a super-organized detective's notebook that brings together three different stories that were previously told separately: the story of the air, the story of the weather, and the story of the traffic. The researchers gathered data from the city's official open data portal, digging through records that go back as far as 2001 for air quality, 2015 for traffic, and 2019 for weather.

The magic of METRAQ isn't just that it has a lot of data; it's how it fits the pieces together. The team took millions of measurements and forced them to speak the same language. They aligned everything to the exact same hour and the exact same location. Imagine a grid of thousands of tiny sensors scattered across the city. At every single hour of the day, METRAQ tells you exactly what the air smelled like, what the weather was doing, and how many cars were zooming by at that specific spot. The result is a dataset that covers 72 months (six years) where all three stories overlap perfectly.

Inside this digital library, you can find 14 different types of air pollutants (like the invisible gases from car exhaust and the tiny dust particles from tires), 7 different weather variables (like wind speed, temperature, and how much it rained), and traffic information derived from over 5,000 sensors. The researchers didn't just copy-paste; they cleaned the data, checked for errors, and even filled in the missing gaps. If a sensor broke for an hour, they used math to guess what the value likely was, marking those guesses clearly so no one gets confused. They even figured out how to estimate traffic levels at spots where there were no traffic sensors, using the data from nearby sensors to paint a complete picture.

The paper doesn't just say, "Here is the data." It also proves that the data works. The authors ran several tests to see if this new dataset could help computers learn to predict the future. They tried to guess what the air quality would be tomorrow based on today's data, and they tried to guess what the air quality was at a specific spot when the sensor there was broken. In these tests, the dataset performed well, showing that it is a reliable tool for scientists and computer programs.

So, what is the big takeaway? The paper suggests that by combining air, weather, and traffic data into one synchronized, high-quality package, we can finally study urban air quality with much greater precision. It doesn't claim to have solved air pollution forever, but it provides the best possible map and compass for anyone trying to navigate the complex relationship between our cars, our weather, and the air we breathe. It turns a messy pile of separate notes into a single, coherent story that can help us build cleaner, healthier cities.

Technical Summary: METRAQ – An Air Quality, Meteorological and Traffic Spatio-Temporal Annotated Dataset

Problem Statement
Urban air quality is a complex phenomenon driven by the interplay of air pollution, meteorological conditions, and traffic activity. While individual datasets for these domains exist, there is a lack of integrated, long-term, spatio-temporal datasets that align these heterogeneous sources at a common resolution and location. Existing datasets often suffer from limited temporal coverage, lack of traffic data, or coarse temporal resolution (e.g., daily aggregates), hindering the development of robust models for urban air quality assessment, exposure analysis, and spatio-temporal forecasting.

Methodology
The authors present METRAQ, a curated dataset constructed via an Extract–Load–Transform (ELT) workflow using open data from the City Council of Madrid, Spain. The methodology involves three primary stages:

Data Extraction and Loading:
- Sources: Air quality data (hourly, since 2001), meteorological data (hourly, since 2019), and traffic data (15-minute intervals, since 2015) were extracted from the Madrid Open Data Portal.
- Normalization: Raw daily-wide CSV files (containing 24 hourly fields per row) were transformed into a "long" format where each row represents a single observation indexed by station, magnitude, and timestamp.
- Validity Preservation: Original validity flags were preserved. A secondary quality control step applied physical plausibility bounds (e.g., $T \ge -55^\circ C$ , $AP \ge 700$ mb) to flag physically impossible values as invalid.
Data Imputation and Augmentation:
- Temporal Imputation: For air quality and meteorological series, missing hourly values were identified. If a station-magnitude-year series had at least 50% valid data, missing values were filled using bidirectional linear interpolation.
- Spatial Interpolation (Meteorology): Meteorological variables were assigned to Air Quality Monitoring Station (AQMS) locations. If a station was co-located, values were copied; otherwise, values were estimated using Ordinary Kriging (if $\ge 12$ stations), Inverse Distance Weighting (if $\ge 4$ stations), or nearest-neighbor interpolation.
- Traffic Augmentation: Traffic data from 5,027 sensors were aggregated to hourly resolution and spatially interpolated to AQMS locations. Five interpolation variants were generated for three traffic variables (Intensity, Occupancy, Average Speed): Radial Basis Functions (Multiquadric, Gaussian, Linear), Inverse Distance Weighting, and Ordinary Kriging.
Dataset Assembly:
- The final dataset consists of 24 annual CSV files (2001–2024) containing 64,881,744 rows.
- It includes 36 magnitudes: 14 air quality pollutants, 7 meteorological variables, and 15 traffic-derived variables (3 base variables $\times$ 5 interpolation methods).
- A flag (is_interpolated) distinguishes between observed and generated values.

Key Contributions
The paper claims four main contributions:

Long-term Integration: A unified dataset covering air quality, meteorology, and traffic for Madrid, with a temporal intersection of 72 months (2019–2024) and air quality data extending back to 2001.
Station-Centered Structure: An hourly resolution dataset aligning 14 pollutants, 7 meteorological parameters, and traffic variables derived from over 5,000 sensors to specific AQMS locations.
Temporal Intersection: A 6-year overlap (72 months) of all three data sources, facilitating joint analysis of urban air quality and mobility.
Technical Validation: The authors provide baseline technical validation workflows for multivariate forecasting, temporal imputation, and spatial imputation to demonstrate dataset usability and reproducibility.

Results and Technical Validation
The paper does not propose new predictive models but validates the dataset's utility through baseline experiments:

Forecasting: Using sensor 28079008 and 8 pollutants (2010–2024), the dataset was tested with models including iTransformer, TimeMixer, TimeXer, MLP, and XGBoost. Results (MSE/MAE) are provided for prediction lengths of 96, 192, 336, and 720 hours.
Temporal Imputation: Using the same setup, missing rates of 12.5% to 50% were introduced. Methods tested included TimesNet, Transformer, linear interpolation, and Autoformer. TimesNet generally outperformed others across missing rates.
Spatial Imputation: Using NO2 data from 2019, the value at a held-out station (28079035) was estimated using data from other stations. Inverse Distance Weighting (IDW) achieved the lowest MSE (0.147) compared to RBF and Kriging methods.
Decomposition: The paper includes MSTL (Multiple Seasonal-Trend decomposition using LOESS) analysis of daily averaged pollutant series, revealing long-term trends and weekly/yearly seasonality components.

Significance
The authors position METRAQ as a resource to support studies in urban air quality, exposure assessment, traffic-related pollution, and spatio-temporal modeling in dense urban settings. By harmonizing heterogeneous public data sources into a common temporal resolution and station-centered structure, the dataset aims to fill the gap between "broad" multi-city datasets and "deep" single-city datasets with limited variable integration. The inclusion of validity flags and interpolation indicators allows researchers to distinguish between observed and derived data, supporting rigorous reproducibility. The dataset is released under a CC-BY 4.0 license with accompanying code for PyTorch integration.

Technical Summary: METRAQ – An Air Quality, Meteorological and Traffic Spatio-Temporal Annotated Dataset

More like this