A strategy for the optimization of quality control checks available in the titanlib open library

<p>Titanlib is a library of functions for the automatic quality control of meteorological observations and it is publicly available on github:</p><p><em>https://github.com/metno/titanlib</em></p><p>Titanlib builds upon the experience of running TITAN (B&#229;serud et al., 2020) for the quality control of temperature and precipitation serving operational applications, such as the production of automatic weather forecasts (Yr.no) and observational gridded datasets (seNorge.no).</p><p>We aim to manage&#160; titanlib as an open project that will eventually incorporate contributions from different working groups.</p><p>The distinctive feature of titanlib is the use of spatial quality control methods. A wide range of spatial checks are made available, from buddy-checks among neighbouring observations to spatial consistency tests (SCTs) based on more sophisticated statistical interpolation.</p><p>This presentation focuses on tuning the parameters of a generic titanlib function. In particular, a general approach to this problem will be presented. A set of reference observations is identified among all available observations. The reference observations are considered good observations. For the sake of simplicity, we call good observations those that are not affected by gross measurement errors, and bad observations those with gross errors. Synop stations constitute the natural reference set; however, other choices are possible. The quality control routine is always applied to the whole set of observations and the statistics are collected only over the reference observations. First, the routine is applied in the &#8220;unperturbed&#8221; mode and we count the number of: correct negatives (i.e. reference observations flagged as good ones), and false alarms (i.e. reference observations flagged as bad ones). Secondly the routine is applied on &#8220;perturbed&#8221; reference observations, by introducing known errors, and we count: misses (i.e. perturbed observations flagged as good ones), and hits (i.e. perturbed observations flagged as bad ones).</p><p>The combination of &#8220;perturbed&#8221; and &#8220;unperturbed&#8221; experiments for different settings of the QC routine allows us to obtain several contingency tables that we use to determine the optimal combination of parameters in an original way. In order to define the cost function, the user is required to specify the prior knowledge on: the expected probability of having a bad observation in the network, and the relative cost of a false alarm for the application considered.</p><p>The method is described by Alerskans et al. (2022), where it has been applied to QC of hourly temperature measured from crowdsourced observations. Furthermore, we will present an application for QC of hourly precipitation.</p><p>References:</p><ul><li>B&#229;serud, L., Lussana, C., Nipen, T. N., Seierstad, I. A., Oram, L., and Aspelien, T.: TITAN automatic spatial quality control of meteorological in-situ observations, Adv. Sci. Res., 17, 153&#8211;163, https://doi.org/10.5194/asr-17-153-2020, 2020</li> <li>Alerskans, E., Lussana, C., Nipen, T. N., Seierstad, I. A.,: Optimizing spatial quality control for a dense network of meteorological stations, accepted for publication in Journal of Atmospheric and Oceanic Technology (JTECH), 2022</li> </ul>