cs.LGOct 7, 2026
SaveSoftSEEPS improves ML-based precipitation forecasting
Organizations: Fraunhofer Heinrich Hertz Institute Berlin, Germany · Technische Universität Berlin BIFOLD Berlin, Germany
Abstract
In this paper we have developed a differentiable approximation of the well-known SEEPS score, which we name SoftSEEPS. This allows the training of a Machine Learning model to forecast precipitation directly. We test SoftSEEPS on the IMERG dataset (0.1 degree resolution) by training a decoder for precipitation on the latent space of a pre-trained low-resolution forecasting model. Combining SoftSEEPS and RMSE in a joint objective is possible with marginal trade-offs in either metric.
Figures & tables
Figure 1: Geographic distribution of the annual-mean "dry precipitation"-probability and threshold values between light and heavy precipitation .
Figure 2: Smooth categorization functions : dashed lines use a smoothing parameter , solid lines .
Figure 3: IMERG test-set RMSE (left) and MAE (right) vs. SEEPS. Circle/triangle: MSE-only and SoftSEEPS -only (Table 1 ). Squares: the combined loss MSE SoftSEEPS swept over .
Figure 4: Locationwise improvement from adding SoftSEEPS to the training loss, for SEEPS (left) and RMSE (right), computed as difference between the evaluation metrics. Blue indicates adding SoftSEEPS is better at that location and in that metric; red indicates MSE only is better.
| Setting | RMSE (mm/day) | MAE (mm/day) | SEEPS |
|---|---|---|---|
| Baselines | |||
| Persistence | 9.288 | 2.927 | 0.819 |
| Climatology | 7.492 | 2.853 | 1.098 |
| ERA5 interpolation | 5.689 | 1.951 | 0.612 |
| Arches+Decoder | |||
| MSE | 5.316 | 1.404 | 0.481 |
Table 1: Next-day precipitation forecasting at (IMERG, test year 2020).
Figure 5: Test-set residual ( , top left, three objectives overlaid) and predicted vs. ground-truth precipitation distribution (remaining panels, one per objective), all log-scale.