← Back to all news
arXiv Machine Learning By Shunya Nagashima, Takumi Bannai

Scale-Recursive Rectified Flows for Few-Step Precipitation Ensembles

Read the original on arXiv Machine Learning →

The paper introduces a scale‑recursive rectified flow that first generates broad rainfall patterns and then refines local details, allocating sampling steps based on ensemble variability and prediction error across spatial scales. This approach reduces the number of sampling steps while maintaining ensemble diversity, improving probabilistic accuracy and rain detection in satellite‑to‑radar downscaling over the United States. The method outperforms a non‑recursive flow with more steps, achieving better accuracy in less time.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Machine Learning.

Related stories

arXiv AI
Sep 16

Composable multi-satellite precipitation estimation for evolving observing systems

The paper introduces PRISMA, a generative framework that separates precipitation prior training from sensor-specific constraints, allowing flexible composition of heterogeneous satellite observations without retraining the core model. By integrating FY‑4B/AGRI, GPM/GMI, F16‑F18 SSMIS, and GPM/DPR‑Ka data, PRISMA consistently improves precipitation‑estimation accuracy and outperforms IMERG Final in CRPS and RMSE while maintaining positive Brier skill across thresholds. The framework supports rapid, accurate ensemble precipitation estimates, enhancing satellite‑based monitoring for hydrometeorological hazards.

By Yunfan Yang, Haofei Sun, Xiuyu Sun, Wei Han, Xiaoze Xu, Xingtao Song, Jun Li, Zhiqiu Gao, Wei Huang
More like this →
arXiv Machine Learning
4d ago

Climate downscaling with likelihood functional flow matching

The paper introduces CD-FFM, a functional flow‑matching model that treats high‑resolution climate fields as continuous random fields, allowing queries at any location and resolution. Unlike grid‑based generative downscalers, CD‑FFM uses a pointwise velocity field and a noise prior fitted to the observed spectrum, giving each observation an exact likelihood. Applied to downscaling ERA5 temperature to 1 km over Fennoscandia, CD‑FFM improves the continuous ranked probability score by 27–31 % over regression and 12–18 % over a larger neural operator, while being about fifty times faster than a CorrDiff‑style diffusion model and also effective for precipitation.

By Zhi-Song Liu, Helmi Toropainen, Ammar Kheder, Michael Boy
diffusion
More like this →
arXiv Machine Learning
Sep 25

Lightweight Probabilistic Downscaling from a Deterministic Base Model

The paper introduces lightweight probabilistic downscaling models that build on a modified U‑Net backbone, adapting two recent machine learning techniques from weather forecasting. Using a two‑stage training curriculum—deterministic pretraining followed by probabilistic fine‑tuning—the authors evaluate their models on the CORDEX‑ML‑Bench suite for daily maximum temperature and precipitation in the Alps, New Zealand, and South Africa. The results show that this approach outperforms the current state‑of‑the‑art in RMSE, offering a more computationally efficient method for generating fine‑resolution regional climate data.

By Joseph McLean, Tiffany Vlaar, Sigrid Passano Hellan, Linus Ericsson
benchmarks
More like this →