Assessing observation network design predictions for monitoring Antarctic surface temperature

Abstract

Networks of observations ideally provide adequate sampling of parameters to be monitored for climate and weather forecasting applications. This is a challenge for any network, but is particularly difficult in the harsh environment of the Antarctic continent. We evaluate a network design method providing objective information on station siting for optimal sampling of a variable, here taken to be surface air temperature. The method uses the concept of ensemble sensitivity to predict locations reducing the most total ensemble variance, that is, uncertainty, across the continent. The method is applied to a network of frequently-reporting stations, and validation is performed using results from assimilating station observations. A cost-efficient "offline" data assimilation framework is used to allow testing over a large sample of experiments, including a large number of randomly chosen networks that serve as a null hypothesis. Network design predictions agree well with observed error reductions from assimilation. The important role of stations on the East Antarctic Plateau in monitoring surface air temperature is evident in network design and data assimilation results, followed by stations in West Antarctica and the Ross Ice Shelf region. Antarctic coastal and Peninsula stations are found to provide the smallest information content integrated over the continent. Validation results are also robust to covariance localization, an essential factor for ensemble methods. Optimal networks outperform randomly chosen-networks in all cases, by up to nearly 50%, depending on the size of the network and the covariance localization distance.

Estimating the benefit of Doppler wind lidars for short‐term low‐level wind ensemble forecasts

“…The 1,000-member ensemble applied mitigates the issue of sampling errors in ESA as shown by Necker et al (2020a). If no large ensemble is available, a time-lagged ensemble or a climatology of forecasts (Hakim et al, 2020;Tardif et al, 2021) could allow a sufficient sample size to be obtained. A statistical sampling-error correction could help to mitigate sampling errors further when relying on smaller ensembles (Necker et al, 2020b).…”

“…However, this balancing cannot be achieved if the analysis influence of observations is localized, as the localized covariances of the Kalman gain ($$$ \mathbf{L}\circ \mathbf{B} $$$) cannot be balanced against the nonlocalized $$$ \mathbf{B} $$$ of the sensitivity. Instead, the variance change estimate can take analysis localization into account by splitting the forecast metric into local components and then applying distance‐based localization weights to the covariances of the individual forecast metrics and the initial state (Hakim et al ., 2020; Tardif et al ., 2021). This approach is equivalent to having a nonlocalized sensitivity and a localized analysis if the observations only influence the forecast within the analysis localization, that is, the impact of the observations does not propagate over time.…”

“…In our companion study (Griewank et al ., 2022), we discuss in depth the errors that result from applying the approach of Tardif et al . (2021) when signal propagation is not accounted for appropriately, and estimating signal propagation is extremely difficult in a nonlinear three‐dimensional atmospheric model. Different variables and perturbations at different scales propagate at differing speeds in the atmosphere, both horizontally and vertically.…”

“…Different variables and perturbations at different scales propagate at differing speeds in the atmosphere, both horizontally and vertically. Tardif et al (2021) can safely ignore signal propagation, as their metric is not calculated from the forecast but from the analysis. By applying Tikhonov regularization to calculate the sensitivity explicitly, we can take analysis localization into account without having to make any assumptions on signal propagation.…”

“…ESA has been used in a number of studies to provide guidance for targeted airborne dropsonde observations for tropical cyclones (Torn, 2014; Majumdar, 2016) and more recently in two network design studies for automatic weather stations in Antarctica (Hakim et al ., 2020; Tardif et al ., 2021). The efficiency of the ESA approach is based on the cancellation of the background covariance $$$ \mathbf{B} $$$ matrix determining the analysis influence and the inverse background covariance $$$ \mathbf{B} $$$ matrix in the ensemble sensitivity in the absence of localization.…”