NLTE Comparisons to Zhenglong's Approach (U Wisc)Tue, Oct 30, 2018
Zhenglong Li (UW) has made an AI-based NLTE algorithm for CriS FSR model spectral that is trained by using longwave channels that have Jacobians that are similar to those in the shortwave region where non-LTE is important. Training is required for each different instrument.
He compared his results to those produced by CRTM and found there were biases (ie their NLTE is not working as well), but he did not compare to SARTA, which we do here.
Here we use the full day of SNPP CrIS observations for 2017/07/17 and compute non-LTE using SARTA and compare those data to Li's calculations.
Figure 1 shows SARTA calcs and CriS obs. Note we matched to ERA and did cloudy calcs, but since these are high altitude channels, we can safely use the cloudy calcs without worrying about cloud temporal mismatches. The 0000:1200 data is used, since this gives an ice mix of ascending (day) and descending (night) plots in one graph. The four panels are
UL : BT2336 obs - the nighttime obs (green, left side) are cooler than the daytime obs (right side, reddish)
UR : BT2336 - BT 667 obs : during night there is no NLTE and so the 667 and 2335 cm-1 channels should see roughly the same BT. This is seen in the descending part of the plot (left) – the difference is close to 0. Conversely there are daytime NLTE effects and so the ascending part (right) sees almost 10 K difference between the two channels
LL : BT667 CrisObs-SARTA cal : close to 0 on average, though we see larger biases in the southern polar region
LR : BT2336 CrisObs-SARTA cal : close to 0 on average, though we see similar larger biases to 667 cm-1 channel in the southern polar region
Figure 1: Top left: BT2336 cm-1 Obs. Top right: BT2336-BT667 obs Bottom left:BT667 cm-1 Obs-SartaCal Bottom right: BT2336 Obs-SartaCal
Top panel of Figure 2 shows a comparison of Zhenglong's biases (using his AI algorithm and his Cris Obs). The bottom panel shows our SARTA biases. Z. Li does have smaller biases, but
- his algorithm needs to be retrained each season
- ours can be improved (a recent development in the RTTOVS NLTE model
uses an additional layer)
Figure 2: Top: Li's results , Bottom: SARTA bias versus ECMWF (obs-calc) B(T).
According to Z. Li, the caption in his plot "Calc-Prediction" should be equivalent to our "Obs-Sarta Calc" as explained below :
"Normally obs - calc_NLTE = obs - (calc_LTE + NLTE). In your case (CRTM as well), NLTE is simulated separately. In my case, NLTE is observation - prediction. So it ends up as prediction - calc_LTE."
Figure 3: Biases using CRTM for 2336 cm-1
Finally Figure 4 shows biases at 667 cm-1 using CRTM (ie Cris Obs-CRTM calcs), which would be compared to the LL panel of Figure 1 above. Again over most of tropical ocean our bias is greenish (close to 0) while the CRTM bias is more cyan (-1 K)
Figure 4: Biases using CRTM for 667 cm-1
- SARTA and Z. Li are quite similar except in the Southern Polar Ocean/Land
- we can do physical retrievals, while sounds like he has to retrain every season
- our algorithm can be improved in the future