GOES-R AWG Hydrology Algorithm Team: Rainfall Potential June 14, 2011

1. GOES-R AWG Hydrology Algorithm Team: Rainfall PotentialJune 14, 2011 Presented By: Bob Kuligowski NOAA/NESDIS/STAR

2. Outline • Executive Summary • Algorithm Description • ADEB and IV&V Response Summary • Requirements Specification Evolution • Validation Strategy • Validation Results • Summary

3. Executive Summary • This Rainfall Potential Algorithm generates the Option 2 product of predicted rainfall accumulation for the next 3 h on the ABI grid. • The NOAA / NSSL K-Means algorithm has been adapted to produce extrapolation-based forecasts of rainfall from the current and previous GOES-R Rainfall Rate fields. • Version 3 (80%) was delivered in September 2010. Version 5 (100%) delivery has been delayed 12 months in response to GPO guidance. • Preliminary validation of the version 3 algorithm has been performed against 20 days of SEVIRI data from 2005. • Validation against Nimrod radar indicates spec compliance.

4. Algorithm Description

5. Algorithm Description • This Rainfall Potential Algorithm generates the Option 2 product of predicted rainfall accumulation for the next 3 h on the ABI grid using the NOAA / NSSL K-Means algorithm. • Rainfall is extrapolated based on a comparison of current and previous Rainfall Rate imagery • ONLY motion is extrapolated (no growth / decay) • No initiation in an extrapolation-based approach • Three basic algorithm components: • Identify features in rain rate imagery • Determine motion between features in consecutive images • Apply motion vectors to create rainfall nowcasts

6. Motivation for Algorithm Selection • In 2007 the Hydrology AT compared the results from 4 different experimental real-time rainfall forecasting algorithms: • Hydro-Nowcaster (NESDIS / STAR) • K-Means (NOAA / NSSL) • TITAN (UCAR) • MAPLE (McGill University) • MAPLE gave the best results but licensing issues could not be resolved; K-Means (second-base performer) was selected instead.

7. Example Rainfall Potential Output Rainfall Potential from 1500-1800 UTC 8 July 2005 derived from Rainfall Rate fields (retrieved from SEVIRI data) at 1445 and 1500 UTC.

8. ADEB and IV&V Response Summary • ADEB did not recommend advancing to 100%, stating “the algorithm needs more validation and reconsideration of the choice of algorithm.” • As a result, the algorithm development schedule has been delayed by 12 months. • Regarding “reconsideration of the choice of algorithm”, the AT responded that the algorithm selection process was well-documented and that validation has been made difficult by the paucity of validation data. The AT also requested clarification on these comments.

9. ADEB and IV&V Response Summary • Regarding “the algorithm needs more validation”, high-quality rainfall data at time scales of 3 h or less have been extremely difficult to find. Planned solutions: • CHUVA (Brazil) field campaign data • Global Precipitation Climatology Center (GPCC) gauge data (can only be used in-house—visit in summer 2012?) • EUMETSAT Hydrology SAF—reaching out to see if their data (also in-house only) could be used via a visit. • In response to a phone request, the AT also surveyed users from OSPO/SAB, NWS/HPC, NWS/OHD, and NWS/WGRFC and found significant interest in the product and for it to be available at launch.

10. Requirements Specification Evolution 10

11. Validation Strategy

12. Validation Strategy • Since the requirement is for rainfall accumulations of 3 h, radar and short-term gauges are the only available source of data for validation against spec • Ground-based radars: • Nimrod radars in UK and Western Europe—5-km grid composite Sample Nimrod 3-h accumulation

13. Validation Strategy • Derive rainfall potential from rainfall rates retrieved from GOES-R ABI proxy data • Meteosat-8 SEVIRI inputs to rainfall rates: Bands 5 (6.2 µm), 6 (7.3 µm), 7 (8.7 µm), 9 (10.8 µm), and 10 (12.0 µm) • Collocate (in space and time) derived rainfall potential with ground validation rainfall accumulations • 3-hour observed rainfall accumulation derived from Nimrod radar • Generate comparative statistics (satellite rainfall accumulation – ground truth rainfall accumulation) • Accuracy • Precision

14. Validation Results

15. Rainfall Potential Validation CDF’s of (absolute) errors of Rainfall Potential vs.Nimrod radar data (Western Europe only) for the 5th-9th of April, July, and October 2005.

16. Rainfall Potential Validation vs. Spec Validation versus Nimrod radar data (covering Western Europe only) for 15 days of data: 6-9th of April, July, and October 2005:

17. Next Steps • AIT determined that the highly object-oriented C++ K-Means code could not be readily implemented into the prototype framework • Consequently, NSSL has been asked to deliver a detailed ATBD and the algorithm is being re-coded from it • Once this is done, the K-Means parameters will be tuned to further enhance algorithm performance: • Experiment with growth / decay option (currently turned off) • Alter minimum rainfall threshold to reduce effects of light rain on motion estimates • Tune scaling parameters

18. Summary • This Rainfall Potential Algorithm generates the Option 2 product of predicted rainfall accumulation for the next 3 h on the ABI grid. • The NOAA / NSSL K-Means algorithm has been adapted to produce extrapolation-based forecasts of rainfall from the current and previous GOES-R Rainfall Rate fields. • Version 3 (80%) was delivered in September 2010. Version 5 (100%) delivery has been delayed 12 months in response to GPO guidance. • Preliminary validation of the version 3 algorithm has been performed against 20 days of SEVIRI data from 2005. • Validation against Nimrod radar indicates spec compliance.