Climateverse contributes predictor datasets to the Google Mumbai Data Challenge

Fine particulate matter can reach deep into the lungs and enter the bloodstream; the World Health Organization links exposure to cardiovascular and respiratory harm. An hourly PM2.5 reading tells us what a Mumbai monitoring station observed. It does not explain what produced that concentration, how conditions differ across the city, or what will happen next. That gap between measurement and explanation is where forecasting research begins.

The problem behind a forecast

The Google Mumbai Data Challenge on Zindi asks participants to predict hourly concentrations of fine particulate matter, or PM2.5, at four monitoring stations. The target is precise; the processes shaping it are not confined to a single site. A multi-year study of Mumbai observations found substantial seasonal variation and complex relationships between PM2.5, other pollutants, and meteorological conditions.

Past readings can reveal a pattern, but a useful forecast must also ask what information was available before the hour being predicted, what geography it describes, and whether it adds insight beyond the pollution record itself. A coarse regional layer cannot simply be read as a street-level measurement. Nor can a dataset assembled after the fact be treated as if it were available in real time.

What evidence could explain a change in PM2.5 at one station and one hour?

A collection built around the research question

Climateverse is a dataset partner for the challenge. We assembled and reviewed 90 openly available predictor datasets spanning weather, roads and buildings, transport and energy, satellite imagery, land cover, heat, population, and other features of Mumbai’s environment. Each source opens a possible line of inquiry into the conditions surrounding an observation.

Bringing these sources together makes comparison possible across data that rarely sit in the same place. Participants can investigate how changing weather interacts with relatively stable features of the city, and whether those relationships differ across monitoring stations. The collection gives those questions a shared starting point.

View full page ↗

The work before modelling

Finding a promising dataset is only the beginning. We reviewed each source for provenance, temporal coverage, spatial resolution, update frequency, access, and known limitations. Climateverse generated a metadata smart card for every record so that these choices and constraints are visible to the people using it.

An hour, a location, and an update schedule are consequential details. They determine whether a proposed predictor could have been known in advance and whether it describes the place the model is trying to forecast. Making them explicit helps researchers design meaningful comparisons instead of simply accumulating variables.

A question-led route through the evidence

A researcher may know the phenomenon they want to investigate without knowing which catalogue terms will retrieve the right data. Climateverse’s collection-specific AI search bridges that language gap. It lets participants begin with a research question and follow the leads into documented sources, where hypotheses can be tested and revised.

View full page ↗

What the challenge can leave behind

A competition creates a defined question and a common measure of performance. The research value can last longer: a transparent collection of evidence, documented well enough for others to examine, combine, and question. The datasets and smart cards can support new analyses of Mumbai’s air beyond a single forecasting task.

At its best, a forecast prompts more than a number for the next hour. It reveals which parts of a city’s environment we understand, which relationships remain uncertain, and where better measurements are needed. Climateverse’s contribution is to make those questions easier to pursue.