Prediction of Landfall Intensity, Location, and Time of a Tropical Cyclone

Sandeep Kumar, Koushik Biswas, Ashish Kumar PandeyAAAI 2021 · DOI

Presented at AAAI-21, this research introduces a high-precision deep learning approach to one of meteorology's most critical challenges: predicting the exact intensity, location, and timing of a tropical cyclone (TC) landfall. By leveraging Long Short-Term Memory (LSTM) networks, the study demonstrates that state-of-the-art predictions can be achieved using surprisingly short observation windows, providing a vital tool for early warning and disaster management.

The Challenge of Landfall Prediction

Predicting a tropical cyclone’s landfall—the specific moment and point where the storm's eye crosses the coastline—is a race against time. While general track forecasting has improved, landfall prediction remains notoriously difficult due to complex, non-linear interactions between the storm, the ocean, and the atmosphere. Traditional dynamical and ensemble models are computationally expensive and often require long-duration historical data to produce reliable forecasts. This research addresses the need for a more efficient, high-accuracy model that can provide early warnings using as little as 12 to 36 hours of recent observation data.

Methodology

The researchers proposed a dual-architecture approach using Recurrent Neural Networks (RNNs) to handle different aspects of the prediction task:

  1. Model-1 (Intensity and Time): A stacked LSTM network with three layers designed to predict the Maximum Sustained Surface Wind Speed (MSWS) at landfall and the time remaining until landfall.
  2. Model-2 (Location): A stacked Bi-directional LSTM (BiLSTM) network with three layers used to predict the latitude and longitude of the landfall point. The bidirectional nature of this model allows it to better capture spatial dependencies and the curvature of the cyclone's trajectory.
Architectural diagrams for Model-1 (LSTM) and Model-2 (BiLSTM).

Key Technical Features

  • Input Features: The models utilize "Best Track Data" (location, pressure, intensity), Sea Surface Temperature (SST), and derived features such as the distance and direction of movement between consecutive 3-hour recordings.
  • Sliding Window: The model processes a sequence of $T$ data points (representing 12, 18, 24, or 36 hours of history) to output landfall characteristics.
  • Optimization: The team used the Adam optimizer with the Swish activation function for Model-1 and ReLU for Model-2 to maximize predictive accuracy.

Datasets and Evaluation

The study focuses on the North Indian Ocean (NIO), encompassing the Bay of Bengal and the Arabian Sea. The primary dataset consists of Best Track Data (BTD) from the Regional Specialised Meteorological Centre (RSMC) of the India Meteorological Department (IMD), spanning 1982 to 2020, supplemented by SST data from NOAA.

Cyclone trajectories in the North Indian Ocean used for training and testing.

After preprocessing to include only cyclones that made landfall, the final dataset comprised 3,988 recordings from 206 cyclones. The models were rigorously tested against three recent devastating cyclones—Bulbul, Fani, and Gaja—which were excluded from the training set to ensure objective evaluation.

Key Results and Insights

The model's performance was evaluated using Mean Absolute Error (MAE). With just 24 hours of input data ($T=8$), the model achieved:

  • Intensity Error: 4.24 knots (MAE).
  • Time Error: 4.5 hours (MAE).
  • Location Error: A distance error of only 51.7 km.

When compared to official IMD forecasts for the cyclone Bulbul at a 72-hour lead time, the model reduced the timing error from 9.6 hours to 1.77 hours and the location error from 112.5 km to 18.2 km.

Core Findings

  • Efficiency with Short Sequences: High accuracy with only 24 hours of data suggests that a storm's immediate momentum and environmental context are highly predictive of its landfall.
  • Bi-directional Advantage: BiLSTMs proved superior for location prediction, likely because they better interpret the "flow" and curvature of a cyclone's path.
  • Feature Engineering: Explicitly calculating the distance and direction of change between observations provided a significantly clearer signal of the storm's velocity than raw coordinates alone.
  • Robustness: The LSTM approach remained consistently accurate even for cyclones that exhibited abrupt direction changes or "re-curving."

Conclusion and Future Work

This deep learning framework offers a significant advancement in landfall prediction, particularly for regions like the North Indian Ocean where early warning is vital. While the current model is geographically constrained and does not yet account for coastal topography, future work aims to incorporate terrain characteristics and expand the model's application to other global oceanic basins, such as the Atlantic and Pacific.

Summary updated

BibTeX
@inproceedings{kumar2021prediction,
  author = {Sandeep Kumar and Koushik Biswas and Ashish Kumar Pandey},
  title = {{Prediction of Landfall Intensity, Location, and Time of a Tropical Cyclone}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i17.17741},
}
Figure