Class imbalance, where one class has significantly fewer training instances than others, is a well-studied challenge in machine learning. However, research on handling class imbalance in time series classification (TSC) remains limited, and no comprehensive experimental comparison of existing approaches has been conducted. Many standard imbalance-handling techniques rely on similarity measures, but computing similarity between time series is more complex than for tabular data. Elastic distances, which account for temporal misalignment, have proved effective in many time series machine learning tasks. We explore resampling strategies for imbalanced TSC and introduce e-SMOTE, an extension of the widely used SMOTE algorithm that incorporates the move-split-merge elastic distance metric. We construct a benchmark of 76 imbalanced TSC datasets derived from the UCR and time series machine learning (TSML) repositories to evaluate state-of-the-art (SOTA) TSC algorithms under class imbalance. Our results show that e-SMOTE enhances the performance of TSC classifiers that typically struggle with imbalance and outperforms both generic and time series-specific rebalancing strategies when tested on our new imbalanced dataset archive.