The Burstinator: A Random Forest for Downburst Detection Using Radar and Lightning Data

Ackermann, Florian; Aregger, Martin; Nerini, Daniele; Balmelli, Simone; Hering, Alessandro; Gehring, Josué; Feldmann, Monika

doi:10.5194/egusphere-2026-4243

Preprints

https://doi.org/10.5194/egusphere-2026-4243

Preprints

30 Sep 2026

| 30 Sep 2026

Status: this preprint is open for discussion and under review for Atmospheric Measurement Techniques (AMT).

The Burstinator: A Random Forest for Downburst Detection Using Radar and Lightning Data

Florian Ackermann, Martin Aregger, Daniele Nerini, Simone Balmelli, Alessandro Hering, Josué Gehring, and Monika Feldmann

Abstract. Operational weather forecasting in Switzerland currently lacks targeted diagnostic tools for severe convective gusts, leaving their occurrence and detailed physical properties insufficiently quantified. In this study, we perform case studies of wind gust-producing convective cells to link 3D C-band polarimetric radar signatures and lightning with gust occurrence. We find that reflectivity cores, K_DP cores, and mid-altitude radial convergence are linked to surface wind gusts. These associations informed the development of the Burstinator, a Random Forest machine-learning model designed to distinguish between thunderstorms with and without severe gusts. The Burstinator yields a 0.66 detection probability paired with a false alarm rate of 0.08, demonstrating superior skill over the baseline WDRAFT. Predictor importance analysis indicates that radar-based features are the top predictors. Finally, we apply the Burstinator to a two-year dataset of 5-min resolution radar data to characterize the spatial distribution of the convective gust frequency. Our findings highlight how integrating machine learning with dual polarization radar and lightning observations can advance convective wind detection and severe weather monitoring across Switzerland.

Received: 17 Jul 2026 – Discussion started: 30 Sep 2026

Competing interests: The position of M.F. at the University of Bern was funded by the Mobiliar Insurance Group. This had no influence any part of this study.

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Florian Ackermann, Martin Aregger, Daniele Nerini, Simone Balmelli, Alessandro Hering, Josué Gehring, and Monika Feldmann

Status: open (until 04 Dec 2026)

Post a comment Subscribe to comment alert

RC1:
'Comment on egusphere-2026-4243', Anonymous Referee #1, 27 Oct 2026 reply
General Comments: The authors present a new machine learning tool that shows skill in distinguishing between downbursts that are associated with severe wind gusts and those that are not associated with severe gusts. The POD, FAR, and CSI are improved compared to a baseline provided by WDRAFT. Predictor importance generated by the random forest algorithm is also discussed, and reflectivity cores, K_DP cores, and mid-altitude radial convergence are the most important predictors. This work is consistent with previous literature, presents a new and potentially useful tool for anticipating downburst impacts, and uses high-quality figures to communicate the results. Overall, the work is very good. There are a few areas that require improvement before publication, especially with respect to expanding (i.e., “filling out”) the results section, adding information and being clearer with some of the methods, and reorganizing the paper to eliminate unneeded/repetitive sections and better pair the results and discussion/interpretation of those results. Therefore, I recommend the paper be accepted after some major revisions. Please see specific and technical correction comments below.
Specific Comments
My largest recommendation for increasing the quality of this paper is adding more results, discussion, and relevance of the results to section 4.2, especially the content related to Fig. 11 (line 238–245), and to a lesser extent Fig. 10 (line 228–237). Currently, line 228–237 and especially line 238–245 read as a somewhat expanded figure caption. It would be valuable to expand on what each figure means for the research and operational community and what the authors think the impact of the results could be. For example, what are the key takeaways about the Burstinator and the predictor values shown by Fig. 11? What might this mean for predictability? What are some potential ways to use/apply these results operationally? What do the random forest results say about the potential usefulness and/or operational application of downburst precursor signatures such as K_DP cores?

Similarly to specific comment 1, the current discussion section (section 5) has some repeated content to section 4 and is not a strong stand-alone section. I find that it is frequently nicer as a reader to have the results and discussion paired together rather than separated into two disconnected sections. Otherwise, the reader is frequently left wondering “so what?” until later in the paper. I recommend removing section 5 and incorporating any relevant content into the appropriate places in section 4. For example, line 289–301 has some similar content to line 220–225. Some of the repeated content could be removed, and the relevant discussion could be moved up to around line 220–225. The same idea applies to line 302–306, which has some similar content to line 247–255.

A few of the methods need further explanation and details, especially with respect to the crowd-sourced data. In general, there should be more information about these data, how prevalent/common the crowd-sourced reports are (i.e., any sense of how often a report comes in when there is actually a downburst), and how they compare with other sources of wind gust magnitude to build confidence that these are good data to use to verify the existence of severe downbursts and the performance of the Burstinator. For example, is there any data about how accurate these crowd-sourced reports might be? Are there any cases/events where a severe gust or EWBD report occurred in populated areas and there was a crowd-sourced report (i.e., when there is a known severe gust in a populated area, is it always accompanied by a crowd-sourced report)? If so, are the crowd-sourced report magnitudes comparable to measured wind gusts (e.g., if a crowd-sourced report is “hurricane”, is a nearby measured gust 30+ m s^-1, or is it only 15 m s^-1)? This information, along with the limitations discussion currently in section 5 (line 271–280) could be added to section 3.1. This idea of accuracy is relevant to the content in line 243–245 where no reports exist but there very well may have been a severe downburst. Trapp and Wheatley (2006; https://doi.org/10.1175/WAF925.1) is probably also a relevant article to cite somewhere. I really like line 134–138 and think it is a great idea, there just needs to be some more information.

Line 125. Why was “strong breeze” rather than “stormy wind” used here? Having reports be only one category apart rather than two might increase confidence that severe/damaging winds really did occur. Please also see specific comment 3.

Line 132. Why might there be a higher proportion of crowd-sourced reports in June and July than August and especially October? Please also see specific comment 3.

Line 19–20, 197–198, and 263. High K_DP can also be indicative of melting hail and/or graupel and melting is likely an important process in increasing the negative buoyancy of wet downbursts. Recommend including this idea and adding Srivastava (1987; https://doi.org/10.1175/1520-0469(1987)044<1752:AMOIDD>2.0.CO;2) and Straka and Anderson (1993; https://doi.org/10.1175/1520-0469(1993)050<1329:NSOMPS>2.0.CO;2) to line 20.

Line 74 and 77. Are there any impacts from using uncorrected differential phase? Were there any clear issues with K_DP and if so, how were those handled?

Line 129. What is meant by “most mature”? Is it the one that has been around the longest? The strongest (e.g., highest Z value)? The tallest? Recommend adding info to clarify.

Line 150. The introduction and interchanging use of radial shear and radial convergence is a little unclear. Recommend using “convergence” whenever talking about the downburst precursor in general (e.g., the first instance on line 150). Using radial shear would be ok when talking specifically about what was done in this study (e.g., second instance on line 150). In addition, something should be added or modified to better connect line 78, line 150, line 194–195, and Fig. 6b.

Line 200. Kuster et al. (2021) does not really include descending K_DP More relevant papers to cite for descending K_DP cores would be Frugis (2020; https://repository.library.noaa.gov/view/noaa/67503) and Gibson and Carlin (2025; https://doi.org/10.1175/WAF-D-24-0222.1). It is possible, if not likely, that the descent of reflectivity and K_DP cores is best seen with rapid-update PAR data (e.g., Heinselman et al. 2008 [https://doi.org/10.1175/2008WAF2007071.1]; Kuster et al. 2016). Seeing descent of these signatures is more challenging with slower radar volume update times (i.e., 5+ minutes).

Line 204. What is meant by “true Gust_Flag”? Is it the same idea as what was introduced in line 121–126? If so, these ideas should be better connected and terminology should either be consistent or defined.

Line 208. What was the magnitude of the crowd-sourced reports (e.g., “Storm”), especially close to the 31.6 and 30.2 m s^-1 measured gust locations? Please also see specific comment 3.

Figure 9 caption. What is meant by “averaged over 100 runs”? What is this 100 runs of? Please reword or add content to make clearer.

Figure 10 caption. What is skexplain? Is it scikit-explain? Some more information should be included about this either in the caption or somewhere in the text (e.g., methods section).

Line 250. Is there is a range bias here? It would make sense that there would be as radar signatures are better sampled closer to the radar, but if such an effect/bias exists, it should be explicitly mentioned somewhere.

Line 256–259. This content also reads like a slightly expanded table caption. Is there any more discussion that could be added here? Please also see specific comment 1.

Line 327. Is it just precipitation loading, or does melting also play a role? I expect melting is important, so precipitation type composition above the melting layer could be important. Updraft strength might also therefore also be important/relevant. Recommend reworking this sentence and perhaps adding content about melting, precip type, updraft strength, etc. that could be supported by your results. Please also see specific comment 1.

Line 342. What is meant by untouched data? Recommend rewording or adding a little bit of content for clarity.

Technical Corrections
Line 33. “Eastern” should not be capitalized.

Line 53, 64, etc. “Random Forest” should not be capitalized.

Line 65. Recommend adding “severe (> 25 m s^-1)” prior to “wind gusts” to be clear what kind of wind gusts are being considered here.

Line 115. Any number range should have an en dash (–) rather than a dash.

Line 198–199. To increase clarity, recommend rewording the sentence to something like “Other case studies, outlined in Ackermann (2025), also showed similar radar properties concerning reflectivity, K_DP, and radial shear.”

Line 200. Recommend changing “Z_DR cores” to “Z_DR columns” to be consistent with previous literature.

Line 258. Recommend deleting “a small fraction of”, so sentence begins as “Only 1% of these TRT cells…”

Section 6. Recommend changing section title from “Conclusions” to “Summary” since the section is a summary of the paper.

Line 324. To increase clarity, recommend rewording sentence to something like “To sum up, the results from the machine learning algorithm helps provide information regarding our research questions as follows:”

Line 341. To increase clarity, recommend replacing “this” with “these data”.

Reply
Citation: https://doi.org/10.5194/egusphere-2026-4243-RC1

Florian Ackermann, Martin Aregger, Daniele Nerini, Simone Balmelli, Alessandro Hering, Josué Gehring, and Monika Feldmann

Model code and software

Burstinator Github Code Florian Ackermann https://github.com/florianvalentinackermann-oss/Burstinator

Florian Ackermann, Martin Aregger, Daniele Nerini, Simone Balmelli, Alessandro Hering, Josué Gehring, and Monika Feldmann

Viewed

Total article views: 592 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
368	181	43	592	45	50

HTML: 368
PDF: 181
XML: 43
Total: 592
BibTeX: 45
EndNote: 50

Views and downloads (calculated since 30 Sep 2026)

Month	HTML	PDF	XML	Total
Sep 2026	59	11	4	74
Oct 2026	287	144	33	464
Nov 2026	22	26	6	54

Cumulative views and downloads (calculated since 30 Sep 2026)

Month	HTML	PDF	XML	Total
Sep 2026	59	11	4	74
Oct 2026	287	144	33	464
Nov 2026	22	26	6	54

Viewed (geographical distribution)

Total article views: 592 (including HTML, PDF, and XML) Thereof 592 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 10 Nov 2026

Short summary

Severe storms in Switzerland cause dangerous winds that are hard to track. We created an algorithm using weather radar and lightning data to detect them. By training it on official weather records and public reports, our algorithm identifies damaging gusts more accurately than current tools. This provides a better way to monitor severe weather, helping protect people and property as these storms may intensify due to a warming climate.


Total:	0
HTML:	0
PDF:	0
XML:	0