Deep learning in cyclone forecasting

Abstract

Prediction of severe weather events such as Tropical Cyclones (TCs) has always been challenging for climate researchers. A disastrous cyclone striking in the coastal region causes serious hazards to human life and economic losses. Intensity predictions are difficult because of the complicated physical mechanisms of TC dynamics and the way they interact with upper-ocean and atmospheric circulation. This research gives importance to estimating TC intensity to identify different categories of cyclones. We attempted to predict TC intensity using Convolutional Neural Networks (CNNs) by proposing a simple and robust architecture for TC intensity estimation. The results yielded better performance than the state-of-the-art techniques with reduced computation time. In addition, we presented a visualisation portal in a production system that displays Deep Learning (DL) output and contextual information for end users. CNN model is trained and tested with classified cyclone data for cyclone identification. The model comprises a binary classifier, a multiclass classifier, a YOLOv3 based cyclone detector and a regression module. The model is tuned for the North Indian Ocean (NIO) region with binary classification accuracy of 98.4% (±0.003), multiclass classification accuracy of 63.83% (±1.3), and Root Mean Square Error (RMSE) of 16.2 (±0.9) knots.