Kandasamy et al. (2025) Hierarchical attention-enhanced multihead CNN and level sets segmentation: A proposed approach to enhance the cyclone intensity estimation

Identification

• Journal: Advances in Space Research
• Year: 2025
• Date: 2025-10-08
• Authors: Lavanya Kandasamy, Ibrahim Ghafir, Sai Harsha Varma Sangaraju, Preksha Mathur, S Rajagopal, Anand Mahendran
• DOI: 10.1016/j.asr.2025.10.001

Research Groups

• School of Computer Science and Engineering, Vellore Institute of Technology, Vellore, India
• Faculty of Engineering and Digital Technologies, University of Bradford, Bradford, United Kingdom
• School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India

Short Summary

This study proposes a novel deep learning approach combining a Hierarchical Attention-Enhanced Multihead Convolutional Neural Network (CNN) with level sets segmentation and Bidirectional Long Short-Term Memory (Bi-LSTM) Networks to improve real-time tropical cyclone intensity estimation and time-series forecasting, demonstrating superior performance over existing methods in the North Indian Ocean region.

Objective

• To enhance tropical cyclone intensity estimation and time-series forecasting by proposing a Nadam-optimizer based Hierarchical Attention-Enhanced Multihead CNN for real-time estimation and Bidirectional Long Short-Term Memory Networks for time series forecasting, incorporating difference of Gaussians and level sets segmentation for pre-processing, which accounts for all physical factors of influence in a cyclone.

Study Configuration

• Spatial Scale: North Indian Ocean region, with geographical coordinates spanning from 60° S to 31.1859° N latitude and 20.0026° E to 146.8982° E longitude.
• Temporal Scale: 2018 to 2023. Infrared (IR) images captured at half-hour intervals, maximum sustained wind speed (MSW) data recorded at 3-hour intervals.

Methodology and Data

• Models used:
  • Proposed: Hierarchical Attention-Enhanced Multihead CNN (with Nadam optimizer), Bidirectional Long Short-Term Memory Networks (Bi-LSTM, with Adam optimizer).
  • Preprocessing: Coordinate-based cropping, Difference of Gaussians (DoG), Level Sets Segmentation.
  • Comparative models (CNN): AlexNet, LeNet, ResNet, VGGNet, CycloneNet.
  • Comparative models (Segmentation): K-Means Segmentation, Mean Shift Segmentation, Watershed.
  • Comparative models (Time Series): GRU, Bidirectional GRU, Attention Bidirectional GRU, LSTM, Stacked LSTM.
• Data sources:
  • Infrared (IR) satellite images: INSAT-3D Imagery (from Meteorological and Oceanographic Satellite Data Archival Centre (MOSDAC), IMAGER-6 channel Level 1 data, IMG_TIR1 Band, BT Dataset Parameter Type).
  • Cyclone maximum sustained wind speed (MSW) details: International Best Track Archive for Climate Stewardship (IBTrACS) - North Indian Ocean dataset.

Main Results

• The proposed Hierarchical Attention-Enhanced Multihead CNN achieved a Mean Absolute Error (MAE) of 28.192 km/h (15.2224 knots) and a Root Mean Squared Error (RMSE) of 36.024 km/h (19.4513 knots) for real-time intensity estimation.
• The model demonstrated superior performance compared to other deep learning models (AlexNet, LeNet, ResNet, VGGNet, CycloneNet), exhibiting lower training and validation losses (13.5701 and 15.5155, respectively) and RMSE values (31.922 km/h and 35.595 km/h, respectively).
• Level Sets Segmentation, as a preprocessing technique, proved most effective, yielding the lowest training and validation losses/RMSE when integrated into the proposed model, outperforming K-Means, Mean Shift, and Watershed segmentation.
• The Bidirectional LSTM model for time-series forecasting achieved a training loss of 20.4799 and a validation loss of 21.2223, with training RMSE of 50.502 km/h (27.2689 knots) and validation RMSE of 45.038 km/h (24.3188 knots), outperforming other time series models like GRU and Stacked LSTM.
• Ablation studies confirmed the critical contribution of each component (attention layer, level sets segmentation, Difference of Gaussians) to the overall model performance.

Contributions

• Introduces a novel deep learning framework for tropical cyclone intensity estimation and forecasting by synergistically combining Hierarchical Attention-Enhanced Multihead CNNs and Bidirectional LSTMs with advanced image segmentation techniques (Difference of Gaussians and Level Sets Segmentation).
• Accounts for physical factors influencing cyclone intensity through sophisticated pre-processing, improving upon traditional Dvorak methods.
• Demonstrates superior performance in both real-time intensity estimation and time-series forecasting compared to several state-of-the-art deep learning models.
• Enhances the interpretability of the predictive model by leveraging CNNs and attention mechanisms to focus on pertinent cyclonic features.
• Provides a robust and effective technique for tropical cyclone intensity estimation, potentially bolstering disaster management and mitigation efforts in the North Indian Ocean region.

Funding

No explicit funding information was provided in the paper.

Citation

@article{Kandasamy2025Hierarchical,
  author = {Kandasamy, Lavanya and Ghafir, Ibrahim and Sangaraju, Sai Harsha Varma and Mathur, Preksha and Rajagopal, S and Mahendran, Anand},
  title = {Hierarchical attention-enhanced multihead CNN and level sets segmentation: A proposed approach to enhance the cyclone intensity estimation},
  journal = {Advances in Space Research},
  year = {2025},
  doi = {10.1016/j.asr.2025.10.001},
  url = {https://doi.org/10.1016/j.asr.2025.10.001}
}