daniel p. tyndall and john d. horel department of atmospheric sciences, university of utah salt lake...
Post on 20-Dec-2015
216 views
TRANSCRIPT
Sensitivity of Surface Temperature Analyses to Background and Observation Errors
Daniel P. Tyndall and John D. HorelDepartment of Atmospheric Sciences, University of UtahSalt Lake City, Utah
Outline
Note: This talk is an excerpt from a paper that has been recently submitted to WAF for review (with M. de Pondeca as a co-author)
Introduction Research motivation Goals
Development of a Local (2D-Var) Surface Analysis Downscaled background Observations Specification of observation error variances and background error covariances
Data denial methodology Hilbert curve withholding technique Root-mean-square error and sensitivity
Case Study Shenandoah Valley, morning surface inversion Results
Summary
Introduction
High resolution mesoscale analyses becoming necessary in variety of fields
Research began in 2006 to help evaluate Real-Time Mesoscale Analysis (RTMA) Estimate error (co)variances of background and observations Identify overfitting problems in analyses
Developed a local surface analysis to help meet these goals Goals of this presentation:
Describe the local surface analysis Present estimates of the background error covariance and
observation error variance Present a data denial methodology to assess analysis accuracy and
identify overfitting problems
Local Surface Analysis (LSA)
2D-Var surface temperature analysis Background
5 km res. downscaled RUC 1-hr forecast RTMA 5-km terrain developed from NDFD
Observations Includes various mesonet and METAR observations ±12 min time window; -30/+12 min time window for RAWS
observations Background and observation errors
Specified in terms of vertical and horizontal spatial distance using decorrelation length scales
Determined using month long sample of observations
Background Downscaling for Temperature
1. Horizontal bilinear interpolation2. Vertical interpolation to height of RTMA terrain
using RUC low level lapse rate RTMA < RUC Elevation: RUC low level lapse rate
multiplied by distance between two elevations and added to RUC 2-m temperature
RTMA > RUC Elevation: RUC 2-m temperature used For complete downscaling description, see
Benjamin et al. 2007 Problem: unphysical features in strong surface
temperature inversions
Observation and Background Error Variances Statistical analysis performed on month-long sample of
observations across CONUS See paper for details; same method used by Myrick and
Horel (2006) Results of analysis show σo
2:σb2 should be doubled (2:1)
Network σb2 (°C2) σb
2 +σo2 (°C2) σo
2 (°C2) Avg. num./hr
ALL 1.4 7.5 6.1 11,464
METAR 2.0 6.2 4.2 1,744
PUBLIC 1.4 6.6 5.3 6,486
RAWS 2.6 12.6 10.0 1,301
OTHER 1.9 8.1 6.2 1,961
Background Error Covariance: Example Correlation - Winchester, VA
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
R = 80 km, Z = 200 mR = 40 km, Z = 100 m
Data Denial Methodology
Evaluation of analyses done by randomly withholding observations
Two error measures: Root-mean-square error (RMSE) calculated at the observation
gridpoints
Root-mean-square sensitivity computed across all gridpoints
Measures need observations that are randomly distributed across the grid to be effective
210
1 1
M Nij ij
j i
a oRMSE
MN
210
1 1
M Lij ij
j i
d cS
ML
Shenandoah Valley Case Study
4°x4° area centered over Shenandoah Valley, VA
Shenandoah Valley between Blue Ridge Mtns. And Appalachian Mtns.
Washington, D.C. located in eastern part of domain
Appal
achi
an M
ount
ains
KIAD
Shena
ndoa
h Va
lley
Blue
Ridge
Mou
ntai
ns0 100 200 300 400 500 600 700 800 900 1000
Washington, D.C.
Case Study: Synoptic Situation
Analyzing analysis generated for 0900 UTC 22 October 2007
Strong surface inversion up to 1500 m in morning sounding
1200 UTC 22 October 2007KIAD
Case Study: Background Field
Downscaling leads southwest-northwest oriented bands
Observations provide detail along mountain slopes and in Shenandoah Valley
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
LSA Analyses
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
R = 40 km, Z = 100 m, σo2/σb
2 = 1 R = 80 km, Z = 200 m, σo2/σb
2 = 2
LSA Analysis Increments
-5 -4 -3 -2 -1 0 1 2 3 4 5
R = 40 km, Z = 100 m, σo2/σb
2 = 1 R = 80 km, Z = 200 m, σo2/σb
2 = 2
Data Denial Example
Data denial methodology applied using 10 observation sets
RMSE and Sensitivity computed for each set of analysis characteristics
Right: Difference between control analysis and data withheld analysis Blue (red) means control
analysis was colder (warmer) than withheld 0 0.5 1 1.5 2 2.5-0.5-1-1.5-2-2.5
Results
# Experiment RMSE Using All Observations (°C)
RMSE Using Withheld
Observations (°C)Sensitivity
(°C)
B Background 2.15 2.15 -
1 R = 40 km, Z = 100 m, σo2/σb
2 = 1 1.62 1.93 0.26
2 R = 80 km, Z = 200 m, σo2/σb
2 = 1 1.80 1.98 0.29
3 R = 20 km, Z = 50 m, σo2/σb
2 = 1 1.41 1.89 0.20
4 R = 40 km, Z = 100 m, σo2/σb
2 = 0.5 1.54 1.94 0.34
5 R = 40 km, Z = 100 m, σo2/σb
2 = 2 1.70 1.93 0.19
6 R = 80 km, Z = 200 m, σo2/σb
2 = 2 1.83 1.89 0.22
7 R = 20 km, Z = 50 m, σo2/σb
2 = 0.5 1.67 1.90 0.26
Measure of analysis quality in data rich areas
Measure of analysis quality in data voids
Summary
Local 2D-Var surface analysis developed for this research
Ratio of observation to background error variance and decorrelation length scales larger than previously assumed
Analysis of RMSE values using withheld observations and all observations provides a measure of analysis overfitting
For further information, see full article submitted to WAF for review
Specification of Observation Error Variance and Background Error Covariance Statistical analysis using month
long sample to estimate error variances See Myrick and Horel 2006
Background error covariance specified in terms of spatial distance:
Estimation shows a σo2:σb
2 of 2:1 and horiz. and vert. decorrelation length scales of 80 km and 200 m
2 22
2 2exp expij ij
ij b
r z
R Z