· 3 min read
Precipitation nowcasting
Nowcasting enables precise short-term mesoscale predictions of weather events. It's crucial in various aspects of public safety, especially during severe, unexpected storms.
Extreme weather events such as precipitation extremes and severe storms cause hundreds of deaths and injuries annually […] Climate change may alter the freqency, timing, intensity and duration of these events. Increases in heavy precipitation have occcurred over the past century. Future climate scenarios show likely increases in the frequency of extreme precipitation events.
— Environmental Health Perspective vol. 109 (May 2001)
Nowcasting forecasts can be generated for any combination of parameters, such as temperature, wind speed, lightning strikes, etc. Our work primarily focuses on the most impactful events such as intense thunderstorms, heavy rainfall, snow and hail by working with real-time radar data of the rain rate (precipitation) in the region.
SAMET System
We have developed the (full form of SAMET here) system that collects all available weather data (radars, sattellites, lightning strikes, ground stations) and exploits the half hourly AI forecast model toissue short-term automatic early warnings. It can also send and receive alerts in the medium term with data computed from the the Italian and European numerical models.
These alerts can be sent to users via various communication channels, including SMS, email, phone calls, and the IO app (pls elaborate on this app). Users can also select which kinds of alerts they want to receive, and specify the areas they want to monitor.
You can try it out here!
Data sources
For this system, we gather data from a large number of sources:
| Precipiation, Wind and Temperature |
|---|
| 187 Regional Stations | 193 Nearby Stations |
|---|---|
| 93 MeteoTN | 55 MeteoSwiss |
| 40 Dams Office | 43 Lombardy |
| 53 Agri | 13 Veneto |
| 23 Crowdsource | 82 South Tryol |
| Lightning Strikes | Radar |
|---|---|
| ITA Airforce | Trentino Radar |
The model
The main model that we use is a radar based precipiation nowcasting deep learning model. It has a spatial resolution of 500 meters and is able to predict upto to 100 minutes into the future with a time step of 5 minutes. It is computationally efficient, taking less than 10 seconds to run.
Our model takes in the previous 30 minutes of data: 
And outputs the predictions (cut short for brevity): 
Mitigating false alarms
As with any deep learning model, false alarms are unavoidable. We have taken two countermeasures to reduce this problem:
AI based anomaly detection
If any measurement deviates too much from its historical behaviour, it is automatically flagged as anomalous and the data source is temporarily disabled.
Alert queue with human review
We have a special, dedicated Telegram channel where all messages are staged for 3 minutes before being sent to all users. The alert is only sent to all distribution channels if no action is taken by the human reviews.
