McCast: Memory-Guided Correction for Precipitation Nowcasting
A new method called McCast addresses error accumulation in long-horizon precipitation nowcasting. Existing autoregressive models drift from physically plausible evolution over long rollouts. McCast introduces a Drift-Corrective Memory Bank (DCBank) that uses temporally organized memory to actively correct latent evolution, rather than passive conditioning. The method is detailed in arXiv:2605.13197v1.
Key facts
- McCast is a memory-guided latent drift correction method for precipitation nowcasting.
- Existing autoregressive methods accumulate errors over long rollouts.
- McCast uses a Drift-Corrective Memory Bank (DCBank).
- DCBank explicitly estimates and corrects drift.
- The method leverages temporally organized memory.
- It actively corrects autoregressive latent evolution.
- The paper is on arXiv with ID 2605.13197v1.
- The approach aims to improve long-horizon forecast accuracy.
Entities
Institutions
- arXiv