RePA: Rebalance Parameter Adaptation for Precipitation Nowcasting

Precipitation nowcasting (PN) is a challenging spatio-temporal prediction task, which aims to predict the radar echo sequences by few historical observations. Existing studies have empirically observed that the prediction result tends to predict the low-echo region, which is hard to capture the details of high-echo parts. Although plentiful works have been devoted to alleviating this problem, little understanding is given to explain it. In this article, we try to give a feasible answer from a long-tailed learning perspective, that is, the bias problem caused by the long-tailed precipitation distribution in natural weather. With this explanation, we reformulate PN as a long-tailed distribution learning problem and solve it by bridging the gaps of the problem between low- and high-echo prediction tasks. To this end, we design a long-tailed distribution learning solution, that decouple and fine-tune the prediction backbone in a rebalanced manner of low- and high-echo region. In particular, we first deeply analyze the long-tailed problem in PM from a structure decoupling perspective and then present two novel decoupling methods, that is, parameter-space decoupling and structure-guided decoupling, for addressing the long-tailed problem. Experimental results on two datasets show that our methods achieve state-of-the-art performance over previous methods.