16C.2 Evaluation of Machine-Learning Based Rapid Intensification Forecast Performance During the 2017 Atlantic Hurricane Season

Friday, 20 April 2018: 11:15 AM
Champions ABC (Sawgrass Marriott)
Alexandria D. Grimes, Mississippi State Univ., Starkville, MS; and A. E. Mercer and K. M. Wood
Manuscript (442.6 kB)

Forecasting rapid intensification (RI) within Atlantic Basin tropical cyclones (TCs) remains a significant challenge in operational meteorology. Recent efforts have addressed this forecast challenge by utilizing an ensemble of artificial intelligence (AI) based prediction models. Development of the AI ensemble required three steps consisting of feature selection, selection of AI ensemble members, and blending of AI ensemble members into a single probabilistic forecast. Data from the Global Forecast System Reforecast database, the National Hurricane Center, and the Statistical Hurricane Intensification Prediction Scheme underwent rigorous feature selection before use as input for RI/non-RI classification. Each AI technique was formulated using numerous configurations, yielding a computationally cheap ensemble of RI predictions from which probabilistic and uncertainty information was derived and individually assessed. Assessment of each method’s performance provided the basis of a blended ensemble for RI prediction. The resulting 41 member AI ensemble was further blended into a single weighted probability of RI for the full AI ensemble suite. Further, as a statistically-based model, robust cross-validation prior to implementation was completed to verify the forecast probabilistic output, contingency and probabilistic forecast statistics. Given the RI/non-RI problem is inherently a probabilistic classification problem, verification statistics were calculated including the probability of detection (POD), false alarm ratio (FAR), Heidke Skill score (HSS), Brier skill score (BSS), and bias. Resulting HSS values were on the order of 0.3 for each AI member and BSS values averaged 0.12 but were as high as 0.38 for the full AI suite. Additionally, for each forecast for the 2017 Atlantic hurricane season, BSS values were evaluated and compared to operational forecast RI probabilities to quantify the improvements offered by the AI-based methods. Results suggest the AI ensemble presents a strong complement to the Statistical Hurricane Intensification Prediction Scheme- Rapid Intensification Index (SHIPS-RII), the current operational forecast product for RI.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner