ConvLSTM-based tropical cyclone intensity estimation and classification using satellite imagery over the North Indian...

Area of Science:

• Meteorology and Atmospheric Science
• Artificial Intelligence
• Remote Sensing

Background:

• Tropical cyclones present significant environmental and societal risks, necessitating accurate identification and intensity estimation for effective disaster prevention.
• Traditional methods for cyclone analysis are often inefficient and lack precision.
• Deep learning offers a promising avenue for advancing tropical cyclone monitoring and forecasting.

Purpose of the Study:

• To develop and evaluate a deep learning framework for automated tropical cyclone detection, classification, and intensity estimation using satellite image sequences.
• To improve the accuracy and efficiency of tropical cyclone analysis compared to traditional methods.
• To explore the potential of hybrid deep learning architectures for capturing spatiotemporal patterns in cyclone data.

Main Methods:

• A hybrid deep learning architecture integrating Convolutional Neural Networks (CNNs) and ConvLSTM was developed to analyze spatiotemporal features in satellite imagery.
• Innovative techniques including clustering-based region isolation, sequence-level data augmentation, and SMOTE for class imbalance were employed.
• Models were trained and validated using the CIMSS Tropical Data Archive and IMD Best-Track datasets with 5-fold cross-validation.

Main Results:

• The VGG16-based model achieved 99.16% accuracy in binary classification of cyclones.
• The ConvLSTM-based model demonstrated 81.1 ± 4.33% accuracy across intensity levels and an RMSE of 7.79 ± 1.27 knots for wind speed prediction.
• The proposed deep learning framework significantly outperformed baseline models in accuracy and predictive capabilities.

Conclusions:

• Deep learning frameworks show significant potential for real-time forecasting and enhancing early warning systems for tropical cyclones.
• The developed hybrid architecture effectively captures complex spatiotemporal dynamics crucial for accurate cyclone analysis.
• Further research involving ensemble learning, advanced architectures, and larger datasets can improve model generalization and forecasting capabilities.