SoftSEEPS improves ML-based precipitation forecasting

2026-10-07

AI Summary

This study addresses the challenge that machine learning-based precipitation forecasting cannot directly optimize the non-differentiable SEEPS score. We propose SoftSEEPS, a novel differentiable approximation of SEEPS, representing its first formulation in this context. Following a differentiable programming paradigm, we construct a deep learning decoder and design a joint loss function combining SoftSEEPS with RMSE, enabling end-to-end training and multi-metric co-optimization on IMERG data. This approach overcomes the gradient backpropagation bottleneck inherent in discrete meteorological scores, making direct optimization of precipitation forecasts feasible. Experimental results demonstrate substantial improvements in forecasting performance, revealing only a negligible trade-off between SoftSEEPS and RMSE. Ultimately, this work establishes a new paradigm for optimizing non-differentiable metrics in meteorological applications.

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.

Problem

• precipitation forecasting
• SEEPS score
• differentiable approximation
• machine learning

Innovation

• Differentiable approximation
• SoftSEEPS
• Precipitation forecasting
• Joint objective
• Machine learning