JEPA · 2026-10-01
PrecipJEPA: JEPA-Regularized Future-State Prediction with Motion-Source Rendering for Precipitation Nowcasting
Yufeng Zhu, Dan Niu, Qiliang Wu, Weiwei Huang, Yixiao Liang, Yongchao Feng, Chunlei Shi
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
PrecipJEPA couples a structured forecasting path with a history-only JEPA auxiliary path for long-term precipitation nowcasting. An online encoder produces spatiotemporal tokens; TFP constructs lead- and location-specific future states; PMSR renders future frames via motion and source–sink fields. During joint training, H-JEPA masks history and predicts latent targets to directly supervise the shared encoder.
Why it matters
The authors address long-term radar-echo evolution while preserving localized high-intensity structures. They argue that encoder history representations trained only from final forecast loss are learned indirectly, and propose H-JEPA to directly shape predictive relationships within observed radar history, paired with time/location-specific future-state construction and motion–source rendering.
Method
- Online encoder maps observed radar sequence to spatiotemporal tokens; TFP constructs lead- and location-specific future states using a dynamics summary and spatiotemporal queries.
- PMSR decodes future states into motion and source–sink fields; transports the latest observation using bilinear warping and adds local intensity changes.
- Joint training uses History-Masked JEPA (H-JEPA): an auxiliary branch masks history and predicts masked latent features via an EMA target encoder, sharing the online encoder with the forecasting path.
Abstract (from arXiv)
Long-term precipitation nowcasting requires modeling radar-echo evolution while preserving localized high-intensity structures. Recent radar-specific studies motivate location-aware prediction and separating echo displacement from intensity change. However existing encoders learn historical representations mainly from final forecast errors. We propose PrecipJEPA, which couples a structured forecasting path with an auxiliary path that enriches its encoder from observed radar history. In the forecasting path, an online encoder first converts the observations into spatiotemporal tokens. The Task-Driven Future-State Predictor (TFP) combines these tokens with a recent-dynamics summary and spatiotemporal queries to construct future radar states. The Parallel Motion-Source Renderer (PMSR) decodes these states into motion and source-sink fields that transform the latest observation into future frames. During joint training, the History-Masked JEPA (H-JEPA) operates on the auxiliary path to predict masked historical features from visible context, directly supervising the same online encoder from the observed sequence. Experiments on SEVIR and MeteoNet show that PrecipJEPA improves highest-threshold CSI by 118.6% and 35.1%, respectively, over the strongest baselines, while maintaining the highest mean CSI throughout the 3-hour forecast.
Related papers
- S2S-JEPA: Predicting the Predictable at Subseasonal-to-Seasonal Timescales
- T-JEPA: A Temporal Joint-Embedding Predictive Architecture for Learning Better Remote Sensing Representations
- Does Joint-Embedding Predictive Architecture Pretraining Help Time Series Forecasting?
- UFMF: Radar-Guided Kinematic Priors for Real-Time 4D Occupancy Forecasting
- TAM: Task-Aware Memory Distillation for Efficient Spatiotemporal Prediction
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-09-30 · Indexed by Pub-AI: 2026-10-01 UTC
PrecipJEPA: JEPA-Regularized Future-State Prediction with Motion-Source Rendering for Precipitation Nowcasting
Preprint posted: 2026-09-30 · Original record ↗
- arxiv
- 2609.38926
Explicit source versions (1)
- v1 · 2026-09-30
- Preprint available
Source date: 2026-09-30 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Long-term precipitation nowcasting requires modeling radar-echo evolution while preserving localized high-intensity structures. Recent radar-specific studies motivate location-aware prediction and separating echo displacement from intensity change. However existing encoders learn historical representations mainly from final forecast errors. We propose PrecipJEPA, which couples a structured forecasting path with an auxiliary path that enriches its encoder from observed radar history. In the forecasting path, an online encoder first converts the observations into spatiotemporal tokens. The Task-Driven Future-State Predictor (TFP) combines these tokens with a recent-dynamics summary and spatiotemporal queries to construct future radar states. The Parallel Motion-Source Renderer (PMSR) decodes these states into motion and source-sink fields that transform the latest observation into future frames. During joint training, the History-Masked JEPA (H-JEPA) operates on the auxiliary path to predict masked historical features from visible context, directly supervising the same online encoder from the observed sequence. Experiments on SEVIR and MeteoNet show that PrecipJEPA improves highest-threshold CSI by 118.6% and 35.1%, respectively, over the strongest baselines, while maintaining the highest mean CSI throughout the 3-hour forecast.
