Assessing the representation of tropical cyclones in ERA5 with the CNRM tracker

Abstract

The ERA5 dataset from the European Center for Medium-Range Weather Foreceasts is the first global reanalysis to reach a horizontal resolution of 31 km and thus provides a unique opportunity to look at tropical cyclones (TC), and in particular at the 3D fields associated with observed TCs. To that end, a specifically calibrated TC tracking scheme is applied on ERA5 along with a track pairing algorithm to match the detected tracks with the IBTrACS catalog in order to investigate how well TCs are represented in the reanalysis. After tuning of the tracking scheme and the application of a dynamic mid-latitude system filtering technique, it is shown that the majority of IBTrACS TCs are detected in ERA5 and that the amount of false alarms is kept reasonably low in most regions. By comparing detected tracks with their IBTrACS counterparts, it is found that TC intensity is still strongly underestimated in ERA5 but that the minimum sea-level pressure distribution is better represented than maximum wind speed. The comparison between the life cycles from both datasets highlights key differences between ERA5 and the best-track catalog, showing in particular that the delay with which TCs from ERA5 reach their peak intensity compared to IBTrACS increases significantly with real TC intensity increase. Finally, the internal structure of TCs in the reanalysis for each intensity class are analyzed and reveal distinct intensification patterns up to Category 3.

Appendix A: tuning the tracker’s detection thresholds

In this section we present a sensitivity analysis of the tracking scheme to its detection thresholds in order to identify a set of tracking parameters (see Sect. 2.2) that maximizes the efficiency of detection while limiting the amount of false alarms. In such cases where two or more objectives are to be simultaneously optimized, no single solution can be derived as there is a trade-off between the objectives. One can however aim to approach a solution that is said to be Pareto efficient if one of the objectives cannot be improved further without degrading the other (Zitzler and Thiele 1999; Deb 2011).

To find such a solution, we explore the parameters’ space by perturbing the tracking parameters so that they can each take three different values, thus forming a set of 243 combinations of five thresholds, called vectors or solutions. The values of each parameters are presented in Table 3. For each vector, our tracking scheme is applied on both the North Atlantic (NAtl) region—which is the best observed basin—and the South Indian (SInd) basin—which is under Météo-France forecasting responsibility—from season 2008 to 2018. We justify why the conclusions drawn from two basins can be largely be applied to other regions at the end of Section A. We then proceed to pair the detected tracks with IBTrACS, following the methodology from Sect. 2.3 and compute the FAR and POD integrated over the 11 year period.

Table 3 Detection thresholds values tested for the sensitivity analysis

Fig. 9

Objective space representation of each combination of thresholds. The first column shows the results for the NAtl basin and the second column those for the SInd basin. Each of the five rows correspond to a detection criterion. Each point (combination of thresholds, a.k.a vector) is colored according to the value taken by said criterion. In each panel, the red-bordered star denotes the selected vector

Figure 9 presents each combination of thresholds in FAR / POD space for both basins, called the objective space. In the objective space, an ideal solution maximizing the POD while minimizing the FAR would be located in the upper left corner of the plots, i.e. POD=100 % and FAR=0 %. However such a solution does not exist here as FAR and POD appear to be related to each other through a non-linear relationship such that improving the POD generally tends to degrade the FAR, and vice versa. In fact, points located along the leftmost of the scatter plots indicate the best trade-off between POD and FAR based on our sample. Choosing a solution among the ones presented here is therefore a subjective choice expressing a balance between detection efficiency and false alarms rate.

These plots also inform us on how each threshold affects the performance of the tracking scheme in terms of these two metrics. For the VOR, most vectors are stacked on top of each other, indicating that the vorticity threshold has little impact on the tracker’s performance, or that the sampling on this criterion was not selective enough. However, and while the effect cannot be seen visually here, higher VOR value appears to be associated with a lower FAR for vectors with high PODs. Indeed, vectors with a POD greater than 70 % show a 1 % relative decrease in FAR with VOR set at $15\cdot 10^{-5}~\text {s}^{-1}$ compared to $5\cdot 10^{-5}~\text {s}^{-1}$ in the NAtl basin, and a 1.3 % relative decrease for SInd. This tends to show that the vorticity threshold may act as a false alarm filter. However, increasing the VOR too much would inevitably lead to a loss in POD as weaker TCs would fail to meet the criterion. Therefore we considered that the gain from testing with higher values would have been marginal with respect to the cost of conducting new experiments. Moreover, the final set of tracks from the global tracking made with the threshold values from Table 3 show a 0.1 % percentile of the maximum vorticity per matched track of $22\cdot 10^{-5}~\text {s}^{-1}$ which tends to support this claim.

The surface wind speed threshold (RES) on the other hand has a clear impact on performances as it effectively defines an upper bound for the POD. Setting RES to 15 $\text {m}.\text {s}^{-1}$ bounds the POD to 54 % in NAtl and 67.3 % in SInd. Boundaries formed by vectors with RES values set to 10 and 5 $\text {m}.\text {s}^{-1}$ are located further up the POD axis and are capped respectively to 74.8 % and 76.7 % for NAtl and 83 % and 83.6 % for SInd. The sensitivity to the RES parameter shows in fact that the POD saturates at RES=10, with RES=5 vectors offering only marginally better PODs in the NAtl basin, and no apparent change in the SInd basin. Because of the POD gap between RES=10 and RES=15, it is possible that the saturation value is actually located between 10 and 15 $\hbox {m s}^{-1}$.

The temperature anomaly threshold (TANOM) stratifies the FAR as each value taken by this parameter corresponds to a certain FAR range with little overlapping. Reducing the threshold leads to an increased FAR, as it allows the tracker to detect cooler systems. However, because of the link between FAR and POD, an increased FAR generally implies a higher POD, making TANOM=1 vectors prime choices for finding a good performing solution within our set.

As for the last two detection thresholds that define the strength of vertical profiles for respectively the temperature and horizontal wind speed (PT and PW), they act by design as filters for extra-tropical cyclones—counting as false alarms in our methodology—by ensuring the presence of respectively a warmer upper core and stronger near-surface winds. Both of these properties are indeed reversed in extra-tropical cyclones, and linked together by the thermal wind relationship. No distinct pattern applicable to both regions and all vectors emerge from the analysis. However, when considering only the group of vectors with the highest POD in each region, increasing these thresholds tends to deteriorate the FAR with marginal benefit to the POD.

As a result, we chose VOR=15 for its potential to reduce the amount of false alarms, RES=5 for the extended POD upper bound, TANOM=1 because of the advantageous location of these vectors in the objective space for the NAtl basin and selected a solution from the remaining candidates in our set which we felt constituted a satisfying compromise in both regions—leaving us with PT=1 and PW=5. The choice of a 5 $\hbox {m s}^{-1}$ wind speed threshold and its meaning for the CNRM TC Tracking Scheme is further discussed in the Disccusion section (Sect. 4). This particular solution exhibits a 54 % FAR and 72.4 % POD in the NAtl region, and a 41 % FAR and 81.2 % POD in the SInd basin. The important amount of false alarms in the chosen solution, the majority of which being mid-latitude systems, motivated us to then add the VTU post-processing method described in Sect. 2.2 as a mean of improving performances even further. Applying the VTU method on the selected solution brings down the FAR to respectively 15 % and 27 % for NAtl and SInd between 2008 and 2018. This reduction in FAR comes at a cost to the POD in the NAtl basin which is lowered to 58 %. However, the POD in the SInd remains unaffected by the use of the additional filter, which is likely due to the fact that there is little TC activity at mid-latitudes in this region.

Finally, it is worth noting that while our test vectors in the SInd region present generally higher PODs in this experiment (14 points more in average) and are arranged differently in the objective space than in the NAtl region, the rankings of each vector on FAR and POD scales is approximately preserved between both region, as can be seen in Fig. 10 below.

Fig. 10

The left column places each vector POD-wise (upper left) and their respective ranks—from 1 to 243—between both experiments (bottom left). The right column does the same FAR-wise. PODs are ranked in ascending order such that higher POD values have a higher rank and FARs in descending order. The red line on the bottom row panels show the fitted line between NAtl and SInd ranks and exhibit a 0.97 and 0.94 correlation coefficient for POD and FAR, respectively

POD ranks between both basins indeed show a 0.96 correlation coefficient FAR ranks are at 0.98, meaning that any given vector performs about as well in both basins with respect to the other vectors. This gives us confidence in the fact that tuning our algorithm over two basins only can be relevant for all basins.