Contrastive Learning Enhanced Diffusion Model for Improving Tropical Cyclone Intensity Estimation with Test-Time Adaptation

Abstract

Tropical cyclone (TC) intensity estimation from satellite images is the very first and critical step of making TC forecasts, whose SOTA performance is achieved by methods built upon CNN based regression models. Unlike discriminative models trained for specific tasks, generative models on the other hand learns to comprehend data in a more sophisticated way through generation. In this paper, we explore the potential of using generative models to further improve the regression task of TC intensity estimation, distinguished from precedents that aim at classification tasks. Our proposed method ConDiff-RTTA optimizes a TC regression model during test time, by back-propagating the loss of a diffusion model conditioned on the regression outputs. More importantly, by enhancing the diffusion model’s training process with our proposed contrastive loss, the diffusion model is more likely to align diffusion losses with prediction errors of the regression model. This enhancement leads to a better understanding of incorrect conditions which facilitates the adaptation of the regression model. We evaluate our proposed method on a benchmark dataset TCIR, where TCs of the latest two years are used as testing cases. Experimental results show that our proposed method ConDiff-RTTA improves the regression model in overall performance, especially on high intensity tropical cyclones. Our code is publicly avalable at https://github.com/maxmaxcu/ConDiff-RTTA/ .