Extreme Wind Speed Prediction Under Noisy Labels: A Transfer-Learning-Assisted Cooperative Sample Selection Approach

Recently, deep learning (DL) techniques have been extensively applied to extreme weather prediction, which demonstrates their potential to address complex meteorological challenges. However, the success of DL-based weather prediction methods relies heavily on the availability of high-quality labelled training data. Human annotators and automated labelling tools may make mistakes due to limited expert knowledge or systematic errors, which leads to the noisy label problem. To address the noisy label challenge, we propose a novel transfer-learning-assisted cooperative sample selection (TLACSS) approach. A leader-follower cooperative learning strategy is put forward to mitigate the effects of noisy labels. To be specific, a leader network is first obtained based on transfer learning. Then, the leader network is jointly trained with two follower networks with the purpose of reducing the prediction divergence among the three networks. The small-loss criterion is employed to identify clean samples based on the joint loss function. A dynamic selection rate is introduced to automatically control the proportion of small-loss samples determined as clean during each epoch. The leader network, trained exclusively on the selected clean samples, is then utilized for extreme wind speed (EWS) prediction using real-world datasets. Furthermore, explainable artificial intelligence techniques are employed to improve the transparency and interpretability of the proposed TLACSS-based EWS prediction method.