Vaisala Radiosonde Absolute Accuracy Response to NDACC Water Vapor Lidar Calibration Workshop
Larry Miloshevich and David Whiteman
This document addresses the action item from the calibration workshop about characterizing the accuracy of Vaisala radiosonde data. This is a summary of published work by Larry Miloshevich on the accuracy of various Vaisala radiosonde models. The corrections referred to below address some of the measurement errors and improve the accuracy. The discussion focuses on the nighttime accuracy since this is the focus of the NDACC group. The basic thrust of this response is that all Vaisala radiosondes (and others) have significant errors in the uncorrected data.
RS92 – Uncorrected
Nighttime Vaisala RS92 RH measurements contain a calibration bias that depends on height and RH (Fig. 9a below, from Miloshevich et al. ). In the lower troposphere it is a mean moist bias that is 3-10% for conditions of RH>10%, and 10-30% for drier conditions. In the upper troposphere it is a dry bias of 10-23% for most RH conditions. Daytime RS92 measurements also contain a dry bias caused by solar radiation error from solar heating of the RH sensor (Fig. 9b), leading to a total dry bias in the UT of 45-50%. The "total uncertainty" should also include the random error of about ±3% (2σ). In addition, "time-lag error" caused by slow sensor response at low temperatures "smooths" the vertical profile in the UT (see Miloshevich et al. ).
Figure 9. The RS92 mean bias relative to the consensus of the reference sensors (CFH, MWR, and SurTHref) is shown for (a) nighttime soundings, and (b) daytime clear-sky soundings with α=62-70°. Polynomials are fit to the CFH comparisons, and the fit extrapolations (dashed) are constrained by the MWR and SurTHref comparisons. The RS92 bias for any data point is given by interpolating between the curve fits. Fits are valid from the surface to 75 mb (night) and 100 mb (day) (vertical black lines). Lower curves give the standard deviation of the bias, offset to 0% at the bottom of the panel for clarity. Coefficients of the curve fits are given in Table 1.
RS92 – Corrected
An empirical correction that removes the RS92 mean bias as well as sensor time-lag error was applied to the data, with the results shown in Fig. 11. The bias uncertainty of corrected RS92 data is estimated from the uncertainty in the reference measurements and the statistics of the dataset to be ±(4% + 0.5 %RH), meaning 4% of the measured value plus a 0.5 %RH component that is significant for dry conditions (dry in an RH sense). For example, for an individual RS92 the total uncertainty, or the RMS sum of the bias uncertainty and the ±3% random uncertainty, for ambient RH=50% is ±6%, while for RH=10% the total uncertainty is ±10%, and for RH=3% the uncertainty is ±21%. The uncertainty estimate is valid for RH>2%, with an upper altitude limit of 18 km. Recent changes since the published correction may eliminate the need for an altitude constraint.
Figure 11. Effect of the time-lag and empirical bias corrections on the Nighttime CFH/RS92 comparison. The RS92 mean percentage difference from CFH, before and after correction, is shown as a function of pressure in 7 RH intervals (left pair), and as a function of altitude (right pair). The upper curves in the left panels show the RS92 mean bias relative to CFH, and the lower curves show the standard deviation of the bias on the same scale, but offset to 0% at the bottom of the panel for clarity. Horizontal line in right panels is the mean tropopause height and red dots are the individual tropopause estimates, and the shaded region is above the upper limit of validity for the correction.
The RS80-H mean bias relative to CFH is shown in Fig. 4 from Miloshevich et al. , before (left) and after (right) applying corrections. In general it is a moist bias for dry conditions (10-50% bias; up to >100% in the UT), and it is a dry bias of 0-20% for moist conditions. Corrections were applied for sensor time-lag error, a T-dependent calibration error as given by Vaisala, and an empirical mean bias correction derived from the AWEX dataset. The residual mean bias relative to CFH is generally within ±10% for all conditions, and within ±5% for RH>40%. Note that RS80 radiosondes are susceptible to sensor icing if supercooled liquid water or prolonged ice-supersaturation is encountered (so is RS92, to a lesser extent, after the heating cycle has ceased).
Figure 4. The RS80-H mean difference from CFH for the AWEX dataset is shown as a function of RH in 3 temperature intervals (left pair), and as a function of altitude for the individual profiles and the mean (right pair). First panel is the original RS80-H data and second panel is after applying corrections. The 3 temperature intervals correspond to the lower troposphere (LT, T>-20°C, solid), middle troposphere (MT, -20>T>-50°C, dashed), and upper troposphere (UT, T<-50°C, dotted). The lower curves in the left panels show the standard deviation of the bias, offset for clarity such that zero is at the bottom of the panel. Horizontal dashed lines in the left panels indicate the AIRS ±10% goal for retrieval accuracy. Horizontal line in the right panels is the mean tropopause height, and dots are the individual tropopause estimates.
The RS80-A mean bias relative to the NOAA frostpoint hygrometer as a function of temperature is shown in Fig. 6 from Miloshevich et al. (2001), before (top) and after (bottom) applying an empirical mean bias correction derived from this dataset. RS80-A data contain a mean moist bias of 0-10% above -10°C, and a mean dry bias that increases with decreasing temperature to >50% below -60°C. After correction the mean bias is within ±10% throughout the troposphere. The increase in the standard deviation for the corrected data below -30°C results from the lack of a time-lag correction in the UT and absence of an RH-dependence in the empirical bias correction.
Figure 6. Ratio of corresponding RS80-A and hygrometer measurements, with curves showing the mean and standard deviation in 1°C temperature bins, for the original RS80-A measurements (left) and after applying an empirical mean bias correction derived from this dataset (right).