Evaluation of Alpine-Mediterranean precipitation events in convection-permitting regional climate models using a set of tracking algorithms

Abstract

We here inter-compare four different tracking algorithms by applying them onto the precipitation fields of an ensemble of convection-permitting regional climate models (cpRCMs) and on high-resolution observational datasets of precipitation. The domain covers the Alps and the northern Mediterranean and thus we here analyse heavy precipitation events, that are renowned for causing hydrological hazards. In this way, this study is both, an inter-comparison of tracking algorithms as well as an evaluation study of cpRCMs in the Lagrangian frame of reference. The tracker inter-comparison is performed by comparison of two case studies as well as of climatologies of cpRCMs and observations. We find that that all of the trackers produce qualitatively equal results concerning characteristic track properties. This means that, despite of quantitative differences, equivalent scientific conclusions would be drawn. This result suggests that all trackers investigated are reliable analysis tools of atmospheric research. With respect to the model ensemble evaluation, we find an encouraging performance of cpRCMs in comparison to radar-based observations. In particular prominent hotspots of heavy precipitation events are well-reproduced by the models. In general most characteristic properties of precipitation events have positive biases. Assuming the under-catchment of precipitation in observations in a domain of such complex orography, this result is to be expected. Only the mean area of tracks is underestimated, while their duration is overestimated. Mean precipitation rate is estimated well, while maximum precipitation rate is overestimated. Furthermore, geometrical and rain volume are overestimated. We find that models overestimate the occurrence of precipitation events over all mountain chains, whereas over plain terrain in summer precipitation events are seen underestimated. This suggests that, despite the convection-permitting resolution, thermally driven thunderstorms are either not triggered or their dynamics still under-resolved. Eventually we find that biases in the spatio-temporal properties of precipitation events appear reduced when evaluating cpRCMs against Doppler radar-based and rain gauge-adjusted observational datasets of comparable spatial resolution, strengthening their role in evaluation studies.

Tracker descriptions

We here provide detailed descriptions of the tracking algorithms and supplementary material.

1.1 MTD

The Method for Object-Based Diagnostic Evaluation (MODE) time domain tool (MTD) is part of the Model Evaluation Tools (MET, see https://dtcenter.org/community-code/model-evaluation-tools-met). It is developed, maintained and made freely available by the Developmental Testbed Center and used here by ICTP. The toolbox comprises various analysis tools developed for the evaluation of numerical atmospheric models. (Davis et al. 2006a, b) first introduce the basic methodology of MODE and demonstrate the advantages of an object-based evaluation in numerical weather predictions. Later on the tracking of objects in time was added and the capability of MTD in describing the characteristic properties of rain systems in both simulations and observations on a continental scale was explored in (Clark et al. 2014). Finally (Prein et al. 2017a) applied MTD in order to identify mesoscale convective systems in convection-permitting climate simulations of North America.

The algorithm can be summarized as follows:

• 1. the field is smoothed by convolution in space, across a radius (here: 1 grid cell, which means smoothing across 3$\times$3 grid cells) and in time across a number of time steps (here: +-0).

• 2. a precipitation threshold (here: 5 mm h$^{-1}$) is applied and thereafter only grid cells exceeding the threshold are considered

• 3. adjacent cells in space and time are clustered to form objects.

• 4. a minimum volume threshold is applied (here: 100 grid cells), meaning that all object that are too small will be dropped.

The output of the analysis is a set of tracks, that represent precipitation events, along with information about their respective location, scale, intensity and propagation.

The location of a track is given through the geometrical centroid across all grid cells associated with the track in space and time. Due to computational limitations the tracker only processes periods of 10 days at a time. Since we are here looking at events with time scales much shorter than that, we don’t expect the analysis being deteriorated much due to this. The mask for comparison of observations against models was applied after the analysis. All statistical properties presented here are derived from the raw un-smoothed precipitation field, whereas the grid cells associated with the event are identified from the smoothed field. Along the boundaries and 1 smoothing radius inwards the input field is set to zero before applying the smoothing.

1.2 OSIRIS

The precipitating system detection and tracking algorithm used by CNRM is based on the algorithm developed at CNRM (Morel and Senesi 2002a, b) applied in precipitation nowcasting at Meteo-France and for the evaluation of the numerical weather prediction model AROME (Brousseau et al. 2016). It has also been recently used in an evaluation study of CNRM-AROME on Mediterranean Heavy Precipitation Events (Caillaud et al. 2021). The 1-hour accumulated precipitation fields are used as input of the tool and the method can be summarised as follows:

• 1. Smoothing : first, each grid cell is replaced by a weighted average of the 3$\times$3 adjacent grid cells and second, a Gaussian filter is applied with a small standard deviation (0.5) allowing for a slight smoothing;

• 2. Detection of the precipitating systems every hour with a minimum surface of 20 km$^{2}$ and contiguous grid points exceeding several intensity thresholds (here: 5 mm h$^{-1}$);

• 3. Tracking of system trajectories by identifying links between systems at different time steps according to overlapping and correlation conditions. The overlapping condition uses the velocity of the cells calculated between different time steps with a minimum recovery rate of 15$\,\%$. The correlation condition is based on spatial correlation calculation between cells at different time steps present in a research box around the cell, with a minimum correlation of 0.4;

• 4. Minimum volume threshold applied (here: 100 grid cells),

• 5. Diagnostics: each cell is schematized as an ellipse: centre of gravity, length of the minor axis and the major axis, angle and the main characteristics of each trajectory can be calculated (location, duration, area, mean and maximum intensity, velocity, ...). The different characteristics are calculated on the smoothed field.

A further description of the algorithm can be found in (Caillaud et al. 2021).

1.3 Celltrack

The tracker (celltrack) used by KNMI is described in detail in (Lochbihler et al. 2017) and is inspired by the work of (Moseley et al. 2013). By default celltrack does not use prior smoothing of the input field. To make celltrack comparable to the other trackers, the input field was smoothed using a $3\times 3$ box-smoothing. This smoothed input field is used in all subsequent steps. First individual cells above a precipitation threshold ($5\,\hbox {mm}\,\hbox {h}^{-1}$) are diagnosed, not considering a specific minimum area. These cells are subsequently linked into tracks. A six-fold iterative advection correction is implemented using advective velocities derived on a coarse-grained grid. After the linking, various track types can be diagnosed (e.g., single “clean” tracks, mergers, splits, etc) following a specific taxonomic classification (Lochbihler et al. 2017). The optional diagnostics of sub-cell and mainstream detection are not used in this study. The Fortran code is available on GitHub.

1.4 DYMECS

The precipitation system detection and tracking algorithm was originally developed for sub-hourly radar and forecast model precipitation data (Stein et al. 2014). Since then, it has been applied to hourly climate model data with resolutions as coarse as 25km (Crook et al. 2019).

The algorithm is divided into two parts: the detection of objects-of-interest for each image and the tracking of these objects-of-interest between consecutive images. The detection algorithm is based on the “local table method” (Haralick and Shapiro 1992), labelling pixels-of-interest by line-by-line scanning. The tracking component is based on the windowed cross-correlation between consecutive precipitation images (Rinehart and Garvey 1978). Windowed correlations between consecutive images are computed, and velocities between images estimated. Objects identified in the previous image are then moved by those estimated velocities. Areal overlap of objects between the two images are computed. If the object overlapping fraction exceeds 0.6, the overlapping objects are considered part of the same track, with splitting and merging allowed if there are multiple overlapping. If two or more objects can be traced back to a single object in the previous image, a split occurs with the object with higher area overlap retaining the same track identifier (i.e., metatrack) and the other objects labelled as new tracks. For the opposite case of two or more objects from the previous image tracing to a single object in the current image, a merge occurs, and retains the identifier with the track with the highest overlap; other merged objects and their identifiers cease to exist.

Smoothing is not originally part of the algorithm. This is added for this study, using the same Gaussian blurring approach as in (Caillaud et al. 2021). Smoothing is applied to the detection phase only, affecting only pixel labelling without changing the underlying precipitation intensities.

The code is written in MATLAB/OCTAVE and is available from the Met Office upon request.