Summary
The impact of snow DA changes has been investigated for 49r1. Snow depth on the Tibetan Plateau and the Rocky Mountains is reduced by assimilating IMS except on complex orography. RMSE of 2m temperature is reduced by 0.5% for winter and 1% for spring in the Northern Hemisphere. In addition, computational time for the 'snow' task is reduced about 30 seconds at 0 UTC by reducing radius for observation scanning and number of IMS input in 2D-OI.
List of snow DA changes
current system | candidate for 49r1 | |
---|---|---|
IMS mask | based on altitude (>1500m) | based on SDFOR (>200) |
IMS thinning | select 1 from every 36 | select nearest IMS on a gaussian grid of 31km (TL639) |
IMS rejection on IMS mask | No | Yes |
Cap value for snow depth | 1.4m | 3.0m |
RSCALE_Z | 800m | 500m |
SCAN_RAD | 3000km | 300km |
ODB | Improve number of O-A |
IMS mask
In the current operational system, IMS is not assimilated on mountainous areas over 1500m (the same as ERA5). However, Orsolini et al.(2019) found:
- The high temporal correlation coefficient (0.78) between the IMS snow cover and the in situ observations
- ERA-Interim, which assimilates IMS on mountainous areas over 1500m, has smaller biases of snow depth on the Tibetan Plateau than ERA5.
In order to improve the biases, the condition to assimilate IMS has been changed. In the latest experiment, IMS is not assimilated if standard deviation of filtered sub-grid orography (SDFOR) is more than 200. This is based on the idea that IMS snow cover from satellite observations could be less accurate on complex orography. The following figures show IMS masks for the current system and the latest experiment.
IMS mask based on altitude (>1500m) | IMS mask based on SDFOR (>200) |
---|---|
IMS thinning
In the current operational system, a data thinning is applied to IMS by using "bufr_filter" to select 1 from every 36 raw observations. However, it leads to inhomogeneous coverage of IMS, especially in some areas. In order to improve it, new data thinning has been tested by using "bufr_grid_screen_parallel" as same as other satellite observations. The thinning selects nearest observations on a gaussian grid of 31km. As a result, the coverage becomes homogeneous and number of observations is reduced from 251926 to 109223.
Simple thinning (select 1 from every 36) | Updated thinning (select nearest IMS on a gaussian grid of 31km) |
---|---|
Cap value for snow depth analysis
Snow water equivalent (SWE) in ERA5, ERA5-Snow and ERA5-Land has been validated against in situ observations (snow course in Canada) by Chris Derksen (ECCC). He pointed out that ERA5 and ERA5-Snow don't represent SWE more than 500mm (personal communication). We found it is caused by a cap value of 1.4m in snow depth analysis. It has been changed to 3.0m in the latest experiment.
Suspicious increments accumulated on IMS mask
The usage of IMS is a bit tricky in the current operational system. IMS is not used on IMS mask in 2D-OI instead of rejecting IMS directly. So large departures can be calculated on IMS mask and lead to large increments outside IMS mask. I found that suspicious increments can be accumulated on IMS mask by the following processes:
- Positive increments by in situ observation at 6, 12, 18 UTC and propagate to grids on IMS mask
- Negative increments around in situ observation by assimilating IMS at 0 UTC
- Positive increments again by in situ observation at 6, 12, 18 UTC ...
In order to reduce such suspicious increments, the following changes have been applied:
- Reject IMS on IMS mask
- IMS thinning at a resolution of 31km (a bit larger than previous setting)
- Reduce vertical correlation (from green line to red line)
Suspicious increments are accumulated on IMS mask before these modifications but they are reduced by them.
control | before modifications | after modifications |
---|---|---|
Radius for observation scanning in 2D-OI
SCAN_RAD has been changed from 3000km to 300km. This change reduces computational cost for sorting observations.
Experiment list
ID | Type | Cycle | Resolution | Start | End | Description |
---|---|---|---|---|---|---|
hut8 | an | 48r1.0 | TCo399 | 02/12/2020 | 28/02/2021 | control |
hwcu | an | 48r1.0 | TCo399 | 02/12/2020 | 31/08/2021 | control (bit-identical to hut8 until 28th Feb) |
hybs | an | 48r1.0 | TCo399 | 02/12/2020 | 31/08/2021 | include snow DA changes |
hut7 | an | 48r1.0 | TCo399 | 02/06/2020 | 31/08/2020 | control |
hybu | an | 48r1.0 | TCo399 | 02/06/2020 | 31/08/2020 | include snow DA changes |
Results
Links to IVER and scorecard
- Iver
- Scorecard
Results for winter 2020/21 and summer 2020 (against Andrew's control)
The following figures shows snow depth averaged in Feb 2021. Snow depth on the Tibetan Plateau and the Rocky Mountains is reduced by assimilating IMS. On the other hand, snow depth on the other mountains are increased by snow cap change of 3.0m.
control (hut8) | test (hybs) | hybs - hut8 |
---|---|---|
RMSE for 2m temperature against analysis in the NH | RMSE for 2m temperature against observations in the NH |
---|---|
Scorecard for winter | Scorecard for summer |
---|---|
Results of 9-month experiments from Dec 2020 to Aug 2021
Computational cost
- Computational time for the 'snow' task is reduced about 30 seconds (2:00 → 1:30) at 0 UTC by reducing radius for observation scanning and number of IMS input in 2D-OI.
- "bufr_grid_screen_parallel" for IMS thinning takes about 25 seconds, a bit longer than "bufr_filter" (15 seconds).
- Both tasks are not on the critical path, so total computational time is not changed by these changes.