CIRL: A Category-Instance Representation Learning Framework for Tropical Cyclone Intensity Estimation
Tropical Cyclone (TC) intensity estimation is a continuous label classification problem, which aims to build a mapping relationship from TC images to intensities. Due to the similar visual appearance of TCs in adjacent intensities, the discriminative image representation plays an important role in TC intensity estimation. Existing works mainly revolve around the continuity of intensity which may result in a crowded feature distribution and perform poorly at distinguishing the boundaries of categories. In this paper, we focus on jointly learning category-level and instance-level representations from tropical cyclone images. Specially, we propose a general framework containing a CI-extractor and a classifier, inside which the CI-extractor is used to extract an instance-separable and category-discriminative representation between images. Meanwhile, an inter-class distance consistency (IDC) loss is applied on top of the framework which can lead to a more uniform feature distribution. In addition, a non-parameter smoothing algorithm is proposed to aggregate temporal information from the image sequence. Extensive experiments demonstrate that our method, with the result of 7.35 knots at RMSE, outperforms the state-of-the-art TC intensity estimation method on the TCIR dataset.