A Semisupervised Deep Learning Framework for Tropical Cyclone Intensity Estimation

Abstract

Tropical cyclone intensity estimation is important to catastrophic weather forecast. In this paper, it is treated as a classification task, with the intensity categories as class labels. Normally, traditional supervised methods require a large amount of prior knowledge for training. However, in reality, only a small amount of labeled samples can be available. Therefore, this paper proposes a novel semisupervised deep learning framework based on convolutional neural networks (CNNs) for FY-4 multispectral images (MSI). The new model only needs a small set of samples labeled a priori to accurately classify the images and estimate cyclone intensity. Moreover, the model involves an iterative training set update process with a hybrid similarity measurement especially designed for the task. The experiments show that the classification performance of the network is improved during the iterations. Evaluation on the estimated intensity categories indicate that the proposed method is significantly better than several existing methods, including the state-of-the-art cyclone intensity estimation model based on CNN, while small training sets are used.