Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SimCast: Enhancing Precipitation Nowcasting with Short-to-Long Term Knowledge Distillation

Yifang Yin, Shengkai Chen, Yiyao Li, Lu Wang, Ruibing Jin, Wei Cui, Shili Xiang

TL;DR

SimCast tackles the challenge of precipitation nowcasting by introducing a two-stage training pipeline that transfers knowledge from a short-term model to a longer-horizon model using augmented sequences and a weighted MSE loss. The long-term model is non-autoregressive, preserving inference efficiency, and can be further refined within a CasCast diffusion framework to improve perceptual quality. Across SEVIR, HKO-7, and MeteoNet, SimCast achieves state-of-the-art CSI and HSS scores, particularly in heavy rainfall regimes, and ablation studies confirm the benefits of weighting, horizon design, and knowledge distillation. This approach provides a practical, accurate, and scalable solution for real-world radar-based nowcasting with potential applications in disaster management and related fields.

Abstract

Precipitation nowcasting predicts future radar sequences based on current observations, which is a highly challenging task driven by the inherent complexity of the Earth system. Accurate nowcasting is of utmost importance for addressing various societal needs, including disaster management, agriculture, transportation, and energy optimization. As a complementary to existing non-autoregressive nowcasting approaches, we investigate the impact of prediction horizons on nowcasting models and propose SimCast, a novel training pipeline featuring a short-to-long term knowledge distillation technique coupled with a weighted MSE loss to prioritize heavy rainfall regions. Improved nowcasting predictions can be obtained without introducing additional overhead during inference. As SimCast generates deterministic predictions, we further integrate it into a diffusion-based framework named CasCast, leveraging the strengths from probabilistic models to overcome limitations such as blurriness and distribution shift in deterministic outputs. Extensive experimental results on three benchmark datasets validate the effectiveness of the proposed framework, achieving mean CSI scores of 0.452 on SEVIR, 0.474 on HKO-7, and 0.361 on MeteoNet, which outperforms existing approaches by a significant margin.

SimCast: Enhancing Precipitation Nowcasting with Short-to-Long Term Knowledge Distillation

TL;DR

SimCast tackles the challenge of precipitation nowcasting by introducing a two-stage training pipeline that transfers knowledge from a short-term model to a longer-horizon model using augmented sequences and a weighted MSE loss. The long-term model is non-autoregressive, preserving inference efficiency, and can be further refined within a CasCast diffusion framework to improve perceptual quality. Across SEVIR, HKO-7, and MeteoNet, SimCast achieves state-of-the-art CSI and HSS scores, particularly in heavy rainfall regimes, and ablation studies confirm the benefits of weighting, horizon design, and knowledge distillation. This approach provides a practical, accurate, and scalable solution for real-world radar-based nowcasting with potential applications in disaster management and related fields.

Abstract

Precipitation nowcasting predicts future radar sequences based on current observations, which is a highly challenging task driven by the inherent complexity of the Earth system. Accurate nowcasting is of utmost importance for addressing various societal needs, including disaster management, agriculture, transportation, and energy optimization. As a complementary to existing non-autoregressive nowcasting approaches, we investigate the impact of prediction horizons on nowcasting models and propose SimCast, a novel training pipeline featuring a short-to-long term knowledge distillation technique coupled with a weighted MSE loss to prioritize heavy rainfall regions. Improved nowcasting predictions can be obtained without introducing additional overhead during inference. As SimCast generates deterministic predictions, we further integrate it into a diffusion-based framework named CasCast, leveraging the strengths from probabilistic models to overcome limitations such as blurriness and distribution shift in deterministic outputs. Extensive experimental results on three benchmark datasets validate the effectiveness of the proposed framework, achieving mean CSI scores of 0.452 on SEVIR, 0.474 on HKO-7, and 0.361 on MeteoNet, which outperforms existing approaches by a significant margin.

Paper Structure

This paper contains 19 sections, 2 equations, 4 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Our Approach
  4. Precipitation Nowcasting
  5. Model Architecture
  6. Optimization
  7. Inference
  8. Short-to-Long Term Knowledge Distillation
  9. Evaluation
  10. Experimental Setup
  11. Comparison to the State-of-the-arts
  12. Ablation Studies
  13. Model design justification
  14. Impact of prediction horizon
  15. Impact of knowledge distillation
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Illustration of our proposed short-term to long-term knowledge distillation for enhanced precipitation nowcasting.
  • Figure 2: The proposed three-step training strategy with data augmentation and knowledge distillation.
  • Figure 3: CSI, POD, FAR comparison at different lead time for threshold 219 on SEVIR dataset.
  • Figure 4: Qualitative comparison of the predictions generated by SimCast and CasCast(SimCast) at different lead times.