Query: Predicting rainfall over an area

[ankit396]

I am working on a project where we want to predict rainfall over a box of coordinates in an area through live data, let's say 3 hours in advance. Is such a prediction is possible with the package(pysteps) for an Indian city? Lets say Monsoon has arrived or will arrive. Can I predict the exact bounding boxes where the rainfall will occur with the best of accuracy.

[ankit396]

Any update?

[dnerini]

Hi @ankit396, and apologies for the delayed reply.

Yes, pysteps can be used to produce short-term precipitation forecasts for an Indian city, provided that you have suitable real-time precipitation observations, which is typically a sequence of quality-controlled weather-radar composites on a regular cartesian grid.

However, there are some important limitations:

• most pysteps nowcasting methods primarily estimate the motion of precipitation observed in the latest radar images and extrapolate it forward. They are therefore best suited to predicting where existing rainfall areas will move during the next approximately 0-2 hours. Have a read at our publication for some theoretical background https://gmd.copernicus.org/articles/12/4185/2019/ .
• a three-hour forecast is technically possible, but uncertainty generally becomes substantial, especially for convective monsoon rainfall. Storm growth, decay, and the initiation of new cells are difficult to predict from radar extrapolation alone.
• nowcasting methods available in pysteps cannot determine that “the monsoon will arrive” from scratch. Predicting the onset of a larger-scale weather system requires numerical weather prediction, satellite observations, or another meteorological forecasting system.
• given these uncertainties, a more appropriate output would be an ensemble or probability map, for example, the probability that rainfall exceeds 1 mm/h within each grid cell or within your area of interest.

For a three-hour operational product, we would generally recommend combining the radar-based pysteps nowcast with a numerical weather prediction forecast. Pysteps contains functionality for probabilistic ensemble nowcasting and for blending extrapolation-based nowcasts with NWP forecasts.

The feasibility and expected accuracy will depend strongly on:

1. the availability, spatial coverage, resolution, and update frequency of the radar data;
2. the quality control and conversion of radar reflectivity to rainfall rate;
3. whether the radar domain is large enough to observe rainfall systems approaching the city;
4. the type of precipitation: organized systems are generally more predictable than rapidly developing convective cells; and
5. validation against historical observations for the specific region and season.

A useful first step would be to clarify what live data you have available: weather radar, satellite precipitation estimates, rain gauges, NWP forecasts, or some combination of these. With radar images every 5–10 minutes, you could start by running one of the pysteps extrapolation or STEPS ensemble examples and evaluate its skill over historical monsoon cases.