Recent BUFR dropsonde data improved forecasts

Bruce Ingleby
Fernando Prates
Lars Isaksen
Massimo Bonavita


On 4 September 2019, ECMWF started operationally assimilating dropsonde reports in the BUFR format. By coincidence this was during the active phase of Hurricane Dorian. Since this was a storm of considerable interest and also well observed by dropsondes, the period was rerun with a) no dropsonde data (‘NoDrop’), b) only alphanumeric dropsonde data (‘ADrop’), c) BUFR dropsonde data instead of the alphanumeric reports where both are available (‘BDrop’). The results, shown in the figure, cover four tropical cyclones (including Dorian) and six tropical storms. They suggest that BDrop is slightly better than the other experiments, especially for intensity. On average, the central pressure was overestimated, but by less in the BDrop experiment. The sample size is rather small, but because of the intermittent nature of dropsonde availability, sample size is usually an issue. Earlier experiments looking at a six‐week period in September/ October 2018 showed a neutral impact of the BUFR dropsonde data (BDrop vs ADrop, not shown).

Advantages of using BUFR data

The 2018 experiments were also used to test the assimilation of low level height data from the dropsonde reports (almost equivalent to assimilating pressure at mean sea level). This is included in the BDrop results shown and in the operational implementation. Alphanumeric dropsonde reports often only have 10 to 30 levels. Some of the BUFR reports, on the other hand, have high vertical resolution (typically about 200 to 900 levels, depending partly on the height of the drop). Some thinning is applied before the assimilation, so that 25 to 30% of these levels are assimilated. Overall, almost six times as many dropsonde wind levels were assimilated in the BDrop experiment as in the ADrop experiment. The BUFR reports also contain the horizontal position of the dropsonde at each level. This makes it possible to account for downwind drift. For the trial periods, the dropsonde drift was relatively modest (20 to 30 km at most).

Position and intensity errors in tropical cyclone forecasts. The left-hand panel shows the average position error and the right-hand panel the average intensity error in ECMWF high-resolution (HRES) forecasts compared to tropical cyclone advisories, for forecast starting dates from 26 August to 8 September 2019 and two forecasts per day. For NoDrop, the 5% and 95% percentiles of the distribution, estimated using a bootstrap procedure, are given. For the other experiments, the percentile bars are similar and tend to move up and down with the mean. The sample size is about 100 at analysis time and 22 after six days (the numbers decline because the storms dissipate both in reality and in the experiments).

Slow migration

While increasing numbers of high‐ resolution BUFR radiosonde ascents have been assimilated at ECMWF since late 2014, the migration of dropsonde data to BUFR has been relatively slow by comparison. In autumn 2018, real‐time high‐resolution BUFR dropsonde reports became available from some NOAA (US National Oceanic and Atmospheric Administration) flights. In the 2019 hurricane season, high‐resolution BUFR reports also became available from some US Air Force flights and from one flight by the Hong Kong authorities (the profiles from the US flights were also available as alphanumeric reports, but the Hong Kong data were BUFR only). Many tropical cyclones are not sampled by dropsondes at all, and in those cases forecast skill is heavily dependent on the good use of satellite data.