Cyclone Intensity Estimation: A Studyof Neural Architecture Search andTransfer Learning in CNNs

Abstract Tropical cyclones (TCs) are severe weather phenomena that can significantly affect human lives. These events can lead to calamities characterized by strong sustainable winds and enormous waves. We proposed an architecture based on convolutional neural networks (CNNs) to tackle this problem. This method makes use of cyclone infrared images. Using customized FCL architectures, we used transfer learning and fine-tuning on CNN architectures such as VGG16, VGG19, and ResNet50, both with and without data augmentation. Fine-tuning involved 4 layers of VGG16, 8 layers of VGG19, and 12 layers of ResNet50 to capture cyclone features effectively. The CNN models were used to these architectures to extract features, and the resulting feature maps were fed to various combinations of Fully Connected Networks (FCL). The most optimistic results were achieved with the VGG16 + FCL (128 x 64 x 1) architecture through transfer learning, producing a Mean Absolute Error (MAE) of 7.51 kts, Root Mean Square Error (RMSE) of 9.63 kts, and an R2 Score of 0.92. Consequently, we identified this model as the foundational basis for Neural Architecture Search (NAS) to enhance the FCL architecture. The NAS process generated various architectures, among which the VGG16 + FCL (128 x 128 x 1) architecture stood out with notable performance, featuring a Mean Squared Error (MSE) of 6.77 kts, RMSE of 8.88 kts, and an impressive R2-Score of 0.945.