Hurricane Forecasting: A Novel Multimodal Machine Learning Framework

Abstract

This paper describes a novel machine learning (ML) framework for tropical cyclone intensity and track forecasting, combining multiple ML techniques and utilizing diverse data sources. Our multimodal framework, called Hurricast, efficiently combines spatial-temporal data with statistical data by extracting features with deep-learning encoder-decoder architectures and predicting with gradient-boosted trees. We evaluate our models in the North Atlantic and Eastern Pacific basins on 2016-2019 for 24-hour lead time track and intensity forecasts and show they achieve comparable mean absolute error and skill to current operational forecast models while computing in seconds. Furthermore, the inclusion of Hurricast into an operational forecast consensus model could improve over the National Hurricane Center’s official forecast, thus highlighting the complementary properties with existing approaches. In summary, our work demonstrates that utilizing machine learning techniques to combine different data sources can lead to new opportunities in tropical cyclone forecasting.

Predicting Landfall's Location and Time of a Tropical Cyclone Using Reanalysis Data

“…Recently, few studies have dealt with TC formation, track, and intensity prediction problems using reanalysis data [13,14,15]. In [15], reanalysis dataset has been used to forecast typhoon formation forecasting in NA, EP, and WP oceans.…”

“…Their model does not take the temporal aspect into account as they have stacked input from two-time steps t and t − 6 to feed into a CNN. In [14], TC intensity and track prediction task is achieved with a lead time of 24h, using reanalysis data ERA5 [17], historical TC data, and output from operational forecast models for NA and EP ocean basins. They have proposed framework Hurricast (HURR) consisting of seven different models that used different combinations of CNN, GRU, Transformers, and XGBoost models.…”

“…They used data between t to t − 21 hours and capture the spatial and temporal aspects of data, thereby addressing the shortcoming of [13]. As there is no existing work that predicts the landfall characteristics from reanalysis and historical TC data, we will try to compare our model with the models presented in [13,14]. [18,19].…”

“…The dataset contains information about ocean basin, latitude, longitude, Estimated Central Pressure (ECP), Maximum Sustained surface Wind Speed(MSWS), distance to land, distance and direction of TC movement, etc at an interval of 3hours. From these features, we choose latitude, longitude, and distance to land for our study and exclude features like MSSW, ECP, distance, and direction of TC movement, which are used by two related works [13,14]. As we are predicting the landfall's location and time of a TC, our dataset consists of all those cyclones which hit the coastal region, and for all such cyclones, only data corresponding to the time points when a TC was moving over the ocean is taken into account.…”

Holistic Deep Learning

“…Deep Neural Networks (DNNs) have yielded great success in a variety of tasks, including computer vision (He et al, 2015;Simonyan and Zisserman, 2014;Szegedy et al, 2014a;Goodfellow et al, 2016), time series modeling, reinforcement learning (Paskov and Bertsimas, 2021), natural language processing, drug design, weather forecasting (Boussioux et al, 2020), and robotics.…”