[Eoas-seminar] MS Defense - Mutschler

[eoas-seminar at lists.fsu.edu]

Good afternoon,

Please join us for Ian Mutschler’s MS Defense on Tuesday, March 10th from 9-11 AM (EST).

Title: The Development and Implementation of Experimental Machine Learning Guidance to Predict Real-Time Probabilities of Landfall for North Atlantic Tropical Cyclones

Name: Ian Mutschler
Date: March 10th, 9:00 – 11:00 AM
Location: EOA 3067
Major Professor: Dr. Robert Hart
Zoom: https://fsu.zoom.us/j/92503253563<https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Ffsu.zoom.us%2Fj%2F92503253563&data=05%7C02%7Ceoas-seminar%40lists.fsu.edu%7C4d80a09693484215d21a08de7888028d%7Ca36450ebdb0642a78d1b026719f701e3%7C0%7C0%7C639080722706066781%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=QwO5p%2FwP4qinXlzz99ceNcWpBXa7FI%2FzfgKkXcoLzhQ%3D&reserved=0>


Abstract:
    Tropical cyclones (TCs) are among the world’s most impactful natural disasters. TCs that make landfall, defined as the intersection of the TC’s surface center with land (\textit{“Glossary of NHC Terms”, n.d.}), can affect large swaths of coastal and inland regions. The areas at and around the exact landfall point often experience the greatest impacts due to the proximity to the maximum sustained winds and the largest magnitude of storm surge. While National Hurricane Center (NHC) TC track forecasts have improved drastically in recent decades (\textit{Cangialosi \& Martinez, 2025}), real-time guidance on if, when, or where a TC will make landfall is not readily available to local decision-makers or the greater public.

    This study aims to develop and evaluate an experimental supervised machine learning (ML) model that incorporates an ensemble of deterministic global numerical weather model output to produce probabilities of landfall for North Atlantic (NATL) TCs at 12hr intervals up to forecast hour 120. Model output is further subset by producing probabilities of landfall for independent regions within the model at each forecast hour range. A suite of predictors from numerical weather model data retrieved from the Automated Tropical Cyclone Forecast System (ATCF) (\textit{Miller et al., 1990}) was developed and used as input into the ML model. Additionally, historical landfall data was retrieved from the NHC’s Hurricane Database (HURDAT2) (\textit{Landsea \& Franklin, 2013}) and was used to create a landfall climatology predictor for model input alongside the model-based predictors described above. With this, a raw training set was built using the model/climatology-based predictors at each 6hr standardized model initialization time (00UTC, 06UTC, 12UTC, 18UTC) and for each 12hr forecast hour time range up to 120 hours. The training set contains model data for TCs over the years 2004-2024 with a seeded random 80/20 split of TCs to be used for training/testing purposes. A landfall was defined in the training set if HURDAT2’s “record identifier” column showed a landfall occurring within the forecast hour time range (i.e. within 48-60hrs from a given model initialization time) – all other cases were considered "non-landfall" cases.

    Three different ML algorithms with varying levels of complexity were trained on the same dataset for evaluation of model skill: logistic regression (LR), random forest (RF), and eXtreme Gradient Boosting (XGBoost). Different ML models were trained at each 12hr forecast time range to account for the growing forecast error in raw model output as the time from the model initialization is increased. Each algorithm’s skill was inter-compared using both objective evaluation metrics as well as subjectively analyzing model performance on case studies from the testing set and from the 2025 Atlantic hurricane season. Model skill was further investigated by evaluating its performance relative to a probability of landfall based purely on climatology (i.e. the percent of TCs at its initialized position that went on to make landfall in a region). This analysis helped select a “best” model for use in producing probabilities of landfall in Atlantic TCs.

    Results show that the ML model developed in this study can produce reliable region-specific probabilities of landfall that exceeds climatology and subjectively performs well in more complex track forecast scenarios. The utility of this model is maximized for cases where the TC’s forecast track is anomalously uncertain. Such examples include recurving TCs where landfalls are highly dependent and sensitive to angle of approach, TCs with high model spread, or TCs approaching smaller landmasses such as those in the Caribbean. As the model’s output is based primarily on deterministic model runs, users can also investigate how landfall probabilities are evolving with time as models converge/diverge on different solutions. With landfall probabilities produced at 12hr intervals, users can also derive the most likely time period where a TC may make landfall. This can be especially useful for local officials making decisions on when to issue evacuation orders or for forecasters predicting storm surge quantities based on local tidal cycles. The work in this thesis addresses a complex forecasting problem by providing decision-makers with real-time probabilistic guidance of TC landfalls up to 120 hours out, improving the critical issue of forecast lead time for landfalling TCs and the communication of associated risks to the general public.

Best,
Adea

Adea Arrison
Sr. Academic Program Specialist
Department of Earth, Ocean & Atmospheric Science
[cid:image001.png at 01DCAA46.F63D08E0]<https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.eoas.fsu.edu%2F&data=05%7C02%7Ceoas-seminar%40lists.fsu.edu%7C4d80a09693484215d21a08de7888028d%7Ca36450ebdb0642a78d1b026719f701e3%7C0%7C0%7C639080722706113570%7CUnknown%7CTWFpbGZsb3d8eyJFbXB0eU1hcGkiOnRydWUsIlYiOiIwLjAuMDAwMCIsIlAiOiJXaW4zMiIsIkFOIjoiTWFpbCIsIldUIjoyfQ%3D%3D%7C0%7C%7C%7C&sdata=SyIALvWx6sA%2FzfZIUkXVSQoSDYmBoGJFkzArY0lwpuQ%3D&reserved=0>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.fsu.edu/pipermail/eoas-seminar/attachments/20260302/d677f934/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image001.png
Type: image/png
Size: 3433 bytes
Desc: image001.png
URL: <http://lists.fsu.edu/pipermail/eoas-seminar/attachments/20260302/d677f934/attachment-0001.png>