Few shot learning for Korean winter temperature forecasts

Abstract

To address the challenge of limited training samples, this study employs the model-agnostic meta-learning (MAML) algorithm along with domain-knowledge-based data augmentation to predict winter temperatures on the Korean Peninsula. While data augmentation has been achieved by using global climate model simulations, the proposed augmentation is purely based on the observed data by defining the labels using large-scale climate variabilities associated with the Korean winter temperatures. The MAML-applied convolutional neural network (CNN) (referred to as the MAML model) demonstrates superior correlation skills for Korean temperature anomalies compared to a reference model (i.e., the CNN without MAML) and state-of-the-art dynamical forecast models across all target lead months during the boreal winter seasons. Sensitivity experiments show that the domain-knowledge-based data augmentation enhances the forecast skill of the MAML model. Moreover, occlusion sensitivity results reveal that the MAML model better captures the physical precursors that influence Korean winter temperatures, resulting in more accurate predictions.