A global rain map learns from sparse satellite radar
PARDIG: Global Precipitation Estimates Learned from Sparse Satellite Observations · arXiv:2611.13941
In plain words
Rain gauges and ground radars do not cover the whole world, so estimating rain everywhere is difficult. PARDIG is a system that learns from sparse rain measurements made by a satellite radar, along with denser satellite images and atmospheric data.
It uses what it learns to estimate rain across a global grid. In comparisons with independent ground radars in the United States, Europe, and Japan, it had the strongest rain-detection scores at every tested rain threshold.
The research
Gaps in rain measurements make it hard to compare precipitation across regions and over time.
A global estimate trained on satellite radar could provide a consistent source for those comparisons, and the independent radar and station results show how well PARDIG detects rain across the places with ground measurements.
PARDIG learns global precipitation estimates from sparse satellite radar readings and dense atmospheric and satellite data, then produces near-real-time and historical products.
PARDIG learns from satellite radar readings at their original, sparse locations and times, then estimates precipitation across a global grid.
Evaluators compared its hourly estimates with independent radar networks and surface stations, using the Critical Success Index, which counts hits against misses and false alarms.
PARDIG-NRT led at every tested rain threshold in three radar regions, and had a higher score than IMERG Final at 89.9% of 36,678 stations. The comparisons support stronger rain detection across the tested regions and stations, though not at every station.
The ground-based evaluations cover instrumented regions and years, while the historical product has lower annual mean precipitation before 2008 and polar estimates beyond about 67° north and south lack direct GPM radar training observations.
What the evidence shows
- From sparse readings to maps
PARDIG, a global precipitation-estimation system, learns from sparse rain measurements collected by satellite radar and dense atmospheric and satellite inputs.
Its estimates fill a global grid, while independent ground-based measurements provide a check on whether the system detects rain well beyond the satellite radar's sampled locations.
- Radar checks across regions
Evaluators compared hourly rain detection with independent radar composites in the United States, Europe, and Japan, using the Critical Success Index (CSI), which counts hits relative to hits, missed events, and false alarms.
PARDIG-NRT had the highest CSI at every tested rain-rate threshold in all three regions, supporting stronger detection across these tested areas.
- Most stations favor PARDIG
For hourly rain detection at the 1 mm/h threshold, evaluators compared CSI at each of 36,678 quality-controlled stations worldwide from 2018 to 2023.
PARDIG-NRT had a higher CSI than IMERG Final at 89.9% of stations, showing a broad advantage in this comparison while leaving room for locations where it did not lead.
- A higher station average
Across those stations, mean CSI was 0.36 for PARDIG-NRT and 0.28 for IMERG Final. This average summarizes station scores; it does not mean every station favored PARDIG.
The two summaries together show both how often each product led and how the average detection score differed across the tested locations.
- A forecast use, with bounds
The evaluations also extend beyond rain estimates: a WeatherNext 3 precipitation head trained on PARDIG had lower forecast error scores against MRMS radar than heads trained on two other precipitation targets at a six-hour lead.
This suggests PARDIG can be useful as a training target for forecasting, though that comparison does not establish that every aspect of the training setup was the same.
Deep
- Original paper authors
- Guy Shalev, Stephan Rasp, Boris Babenko, Andrew El-Kadi, Mohammed Alewi Hassen
- Submitted to arXiv
- License
- CC BY 4.0