Identifying ice accretion on wind turbine blades is of paramount importance for operations in frigid climates. Accurate detection is crucial to preemptively address potential revenue losses and diminished power generation. While numerous machine learning paradigms have been engineered to enhance the precision of blade icing detection, they frequently encounter difficulties in effectively managing sensor correlation dynamics and the substantial class imbalance inherent in blade icing datasets. This often leads to suboptimal precision and elevated false alarm rates. In this study, we introduce a novel approach to tackle these challenges, thereby fostering more dependable blade icing detection. Our contribution is a Spatial-Temporal Graph Convolutional Network (SGCN), which leverages graph convolutional networks to adaptively analyze the intricate relationships among sensors. Furthermore, we integrate a distance-based classifier to bolster performance in class-imbalanced learning scenarios. Empirical evaluations conducted on both publicly available UEA time-series datasets and real-world wind turbine data substantiate that the SGCN surpasses alternative methodologies, particularly in situations characterized by extreme data imbalance.