Open data and external sources

Before using a source we study its access, formats, licence and actual latency, measured with real calls rather than taken from the documentation. Each source has a row in a licence register, with its status and attribution; we leave out those with incompatible conditions, for example a ban on commercial use. The most important reports go through an adversarial review, with an explicit vocabulary: verified, estimate, assumption, to be verified.

One downloader per source, as a system service: it resumes on its own, skips what it already has, makes one request at a time with the required pauses and respects the sources’ limits. We keep all the raw data, losslessly compressed and with a sha256 hash, so every product can be recomputed. We add space; we do not stop the sources.

From the raw data we derive nationwide fields and tables in documented and versioned formats, each with its source, licence, attribution and three timestamps: the time of the data, the time of publication at the source and the time of processing. Guards are part of the product: an empty response or one with a different shape is an error, not data.

Every 10 minutes we acquire the hotspots from the VIIRS and MODIS satellites and fire radiative power from MTG and Sentinel-3. We separate fires from industrial and volcanic heat, track the fires over time and cross-reference them with wind and land use.

Every night we download the VIIRS images of the whole of Italy and, as soon as they are released, the calibrated Black Marble products; we compare the light of an area with previous clear nights and from this derive events with a start, a duration and a restoration. It is a confirmation, not an alarm on its own.

We continuously collect models, observations and images from official sources, always for the whole of Italy or for fixed bounding boxes, with plausibility checks that flag a suspect value without altering the measurement.

We combine water level measurements, discharge forecasts, flooding observed from satellite and hazard maps to tell where a river is rising, where a flood is forecast and what is exposed. We state the limits of each source next to the data.

We follow INGV earthquakes every minute with all versions of their location, keep every ShakeMap version and immediately compute an estimate, labelled as such, of the felt area and of tsunami arrival times. The estimate gives way to the official products as soon as they are released.

We collect public data on the European electricity system (cross-border flows, balancing, generation by zone and energy source, unavailability of power plants and interconnections), correcting known defects during collection, and we weight them by the population of census sections rather than by average densities.

The whole chain: collection from news services and local feeds, pre-filtering, duplicate detection, classification with AI models into configurable channels, events, geolocation and an explainable criticality score, always broken down into its points. We do the geolocation in-house, without external services, and it is never more precise than the text.

Alerts on three levels (Advisory, Warning, Alarm) with thresholds anchored to technical references, calibrated against the archives and measured for notification load before we activate them: an operator swamped by notifications does not see what matters.

We download the open base geographic datasets (LiDAR surveys, regional orthophotos, buildings with heights) in full and by region, convert them for 2D and 3D maps and measure their quality before using them.