Tropical Cyclone Wind Speed Estimation: A Large Scale Training Data Set and Community Benchmarking

Abstract

BackgroundTropical cyclones (TCs) cause damage to life and infrastructure, exceeding $10 billion (Blake et al., 2011) in losses in extreme cases. There are numerous factors, including rainfall, wind speed of the TC, and storm surge (NOAA, Hurricanes, 2021), that contribute to the amount of damage. However, due to the spatial scale of wind damage and sparse, low frequency observations over water, estimations of the intensity (wind speed) of TCs is difficult. Hence, methods like the Dvorak technique (Dvorak, 1984), Advanced Dvorak technique (Olander & Velden, 2007), and Deep Learning to Estimate Tropical Cyclone Intensity from Satellite Passive Microwave Imagery (Wimmers et al., 2019) have been developed that use more regularly available satellite observations to estimate the wind speed of TCs (Clarizia et al., 2014;Clarizia & Ruf, 2016).The Dvorak technique (Dvorak, 1984) was initially developed in 1969 and involves human inspection of satellite imagery to observe the specific cloud structure of the TC and classify them into appropriate wind speed categories. This intensity estimation method introduces uncertainty in the estimations due to its dependency on the human interpretation which is inherently subjective. This provides a unique opportunity to utilize the power of ML for TC wind speed estimation. For any supervised ML model to perform well, a well curated, sufficiently large labeled data set is necessary (Virts et al., 2020). Moreover, well curated data are highly sought after within the general ML community as it alleviates the requirement of domain expertise for training ML models. For these purposes, we present a sufficiently large, curated TC wind speed estimation data set based on Geostationary Operational Environmental Satellite (GOES) satellite imagery. The data are presented in a way that can be leveraged to model time-series-based predictions as well as single frame-based regression predictions. GOES satellite imagery, with data available every 10 min (Advanced Baseline Imager, n.d.), has higher temporal resolution than advisories issued by the National Hurricane Center (NHC) (NHC Advisories, 2021), which are released every 6 hr. A model based on these observations can potentially reduce the time between wind-based intensity estimations from 6 hr to 10 min. Furthermore, use of ML model will aid the process of categorization, making the process more efficient while also reducing human induced errors or biases. The estimations from these models can also be analyzed to understand the salient features of a tropical cyclone contributing to their wind speed categorization.