Answers are grounded in this paper and cite their sources.
Evaluates self-supervised learning methods for improving typhoon cloud forecasting, indicating superior performance with tailored adaptations.
Self‐supervised learning (SSL) provides an innovative paradigm for pretraining without involving any new data or labels. However, its potential has not yet been evaluated in the field of typhoon raincloud nowcasting. Therefore, we explore two different SSL approaches: autoencoder (AE) and masked autoencoder (MAE). Both are then paired with a subsequent adaptation strategy, namely fine‐tuning (FT). We evaluate the performance gains from different combinations using the popular U‐Net backbone across multiple evaluation metrics. The results show that both AE‐FT and MAE‐FT outperform control experiments (Ctrl), where the U‐Net is trained from scratch without SSL. The AE‐FT yields a root mean square error of 6.56, around 2.6% improvement compared to Ctrl (6.72), whereas the MAE‐FT excels in extreme event detection, achieving a Heidke Skill Score of 0.12 at the 40 dBZ threshold, 50.7% higher than Ctrl (0.08). Comparison with rotation‐based data augmentation further validates the merits of SSL for pretraining, which stem from enrichment of learned representations in general and geospatial sharpness induced by masking strategies. However, the choice of loss function used in SSL greatly impacts its effectiveness, and a perceptually sensitive loss function like structure similarity (SSIM) loss is more desirable than the traditional mean square error (MSE) loss.
No takes yet. Share an insight, caveat, or question.
Yao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: