PaperlayerView exact revision on arXiv

Source-linked AI summary

Lightweight Probabilistic Downscaling from a Deterministic Base Model

Joseph McLean, Tiffany Vlaar, Sigrid Passano Hellan, Linus Ericsson

arXiv:2609.29383v1cs.LG

TL;DR

Regional downscaling can supply finer climate information than coarse global models, but probabilistic forecasting advances have not yet been applied to this task. This paper adapts a two-stage U-Net approach and reports state-of-the-art RMSE, while the costly generative competitor remains stronger on distributional metrics.

  • Problem

    Global climate models are too coarse for many local decisions, and recent probabilistic forecasting advances have not yet benefited climate downscaling.

    Source: §1 · pp. 1–2

  • Method

    The paper adapts a U-Net with deterministic MAE pretraining followed by probabilistic CRPS tuning and MC dropout for regional climate downscaling.

    Source: §3 · pp. 2–3

  • Results

    The models set state-of-the-art RMSE for temperature and precipitation, but the expensive generative competitor far exceeds them on distributional metrics.

    Source: §1 · p. 2, §5 · p. 4

  • Takeaways & Limitations

    The results advance lightweight probabilistic downscaling, though performance on distributional metrics remains behind the fully generative competitor.

    Source: §5 · p. 4

    The models are substantially outperformed by the expensive fully generative competitor on metrics targeting prediction-distribution properties.

    Source: §5 · p. 4

Abstract

from arXiv · show

Climate data downscaling is the task of increasing the spatial resolution of climate data, typically by generating fine-resolution regional climate data from coarse global model output. Recent machine learning (ML) work in the related task of weather forecasting has seen significant improvements due to newly devised training methods and architectural components, but these have not yet benefited downscaling. We adapt two of these methods to create a family of lightweight probabilistic ML downscaling models built on a modified U-Net backbone and evaluate them on the CORDEX-ML-Bench suite for daily maximum temperature and precipitation across three geographic regions: the Alps, New Zealand and South Africa. We find that a two-stage training curriculum, combining deterministic pretraining with probabilistic tuning, transfers well to downscaling, beating the state-of-the-art for RMSE. Our work provides an advancement towards lightweight, probabilistic downscaling models, reducing the current trade-off between computational intensity and distributional fit.

1 Introduction

Downscaling can provide finer regional climate data than global models typically produce, supporting decisions that require local information. This paper adapts efficient probabilistic forecasting methods to downscaling and reports state-of-the-art RMSE on temperature and precipitation.

1 Introduction

  • High-resolution regional temperature and precipitation data support decisions in agriculture, infrastructure resilience, and extreme-weather preparation.Most global climate models operate at resolutions of tens to more than a hundred kilometres, while even HighResMIP experiments are generally limited to 25–50 km.Source: §1 · p. 1
  • Deterministic models can blur predictions and underestimate extremes, while generative models better recover fine-scale statistics at substantially higher training and inference cost.The paper frames this as a trade-off that recent probabilistic forecasting work suggests need not be fundamental.Source: §1 · p. 1
  • The paper adapts a two-stage deterministic-to-probabilistic curriculum and other forecasting advances to downscaling, reporting state-of-the-art RMSE for temperature and precipitation on CORDEX-ML-Bench.The models are described as lightweight enough to train on a single GPU, while further work is needed on distributional metrics.Source: §1 · p. 2

2 Related Work

Prior downscaling methods include deterministic networks that can underestimate extremes and generative methods that better reproduce fine-scale statistics at greater computational cost. Recent forecasting work offers efficient probabilistic approaches, which this paper brings to downscaling.

2 Related Work

  • Deterministic downscaling models commonly estimate targets by mean or absolute error and tend to produce blurry outputs that underestimate extremes.CORDEX-ML-Bench standardized comparisons across 40 ML configurations and three regions, identifying extreme-value underestimation as a consistent failure mode.Source: §2 · p. 2
  • Generative approaches can recover fine-scale statistics, but GANs may train unstably and diffusion models can be expensive and unsuitable for lightweight deployment.CorrDiff uses a deterministic U-Net to predict the mean, then a corrective diffusion model to refine the residual.Source: §2 · p. 2
  • Weather-forecasting methods combine U-Cast's deterministic pretraining and probabilistic fine-tuning with MOSAIC's learned perturbations and block-sparse attention.MOSAIC reports matching state-of-the-art models with an order-of-magnitude speed-up.Source: §2 · p. 2
  • This paper adapts forecasting advances to downscaling, where the model increases spatial resolution rather than predicting the next time step.The authors describe the pipeline as having potential to reduce computational cost and increase accessibility.Source: §2 · p. 2

3 Method

The method adapts a U-Net and a two-stage deterministic-to-probabilistic training curriculum for downscaling. It adds decoder upsampling and investigates alternative stochastic mechanisms and attention components.

3 Method

  • The U-Net is trained first with MAE for deterministic feature learning, then with CRPS and MC dropout to produce prediction ensembles and learn uncertainty.The probabilistic second stage is shorter than deterministic pretraining.Source: §3 · p. 2
  • Three added decoder upsampling layers convert a 16×16 input grid to a 128×128 output for regional downscaling.The task changes from predicting a future weather time step to estimating fine-grained regional climate data from coarse input.Source: §3 · p. 3
  • Learned functional perturbations inject gated Gaussian noise into intermediate features so the network can model distributions through noise samples.The learned gate controls how strongly perturbations affect different parts of the network.Source: §3 · p. 3
  • Block-sparse attention combines coarse global context, selected distant information, and local information without computing attention for every token pair.The method is intended to capture climate relationships efficiently.Source: §3 · p. 3

4 Experiments

Across three regions and two climate variables, the experiments compare model variants with benchmark leaders and examine the two-stage training curriculum. The strongest RMSE results come from MC dropout with block-sparse attention, while distributional metrics remain a challenge.

4 Experiments

  • Models are trained separately for the Alps, New Zealand, and South Africa, using 100 deterministic epochs followed by eight probabilistic epochs for tasmax and precipitation.Evaluation metrics are reported as means across the regions; experiments use a single 24 GB NVIDIA GeForce RTX 3090.Source: §4 · p. 3
  • The evaluation covers RMSE, climatological mean, RALSD, inference memory, and variable-specific metrics for extremes, wet days, interannual variability, and distributional fit.The comparison includes the benchmark's top generative model RCMGEM-mv-orog and deterministic models ParamUNET and DeepESDcrps.Source: §4 · p. 3
  • MC dropout with block-sparse attention achieves the lowest RMSE for tasmax and pr across the benchmark; RCMGEM-mv-orog leads most other metrics.The temperature table identifies both MC-dropout variants as most accurate for RMSE; for precipitation, MC-dropout with block-sparse attention is most accurate for RMSE and LHD.Source: Table 1 · p. 3, Table 2 · p. 4, §4 · p. 4
  • The models achieve the benchmark's lowest RMSE but are much worse than the benchmark's worst model on precipitation SDII.The authors report that their scores are within the benchmark range for all but two metrics, including IAV for tasmax and SDII for pr.Source: §4 · p. 4
  • Two-stage training outperforms training with only stage one or only stage two on overall RMSE; stage-two-only training also increases memory without improving RMSE.The separate stage-two ablation reports RALSD decreasing faster than RMSE at 10 epochs, suggesting longer stage-two training could improve distributional metrics.Source: §4 · p. 4, Table 3 · p. 4

5 Conclusion

The lightweight two-stage models set a new RMSE state of the art, but lag on metrics targeting distributional properties. The conclusion notes both the cost advantage and room for further tuning.

5 Conclusion

  • The models set a new state of the art on RMSE while falling behind on metrics targeting specific statistical properties of the prediction distribution.They improve some of these scores over the deterministic baseline, but the top generative model, though very expensive, exceeds them.Source: §5 · p. 4
  • The paper introduces cheap, lightweight downscaling models trained with a two-stage deterministic-to-probabilistic curriculum.Source: §5 · p. 4
  • The study did not exhaustively tune the U-Cast adaptation, and the authors expect a longer second stage could improve non-RMSE metrics.They identify stage lengths, optimizer, and learning-rate scheduling as possible tuning targets.Source: §5 · p. 4

A Further Results on the Two-Stage Ablation

In a 10-epoch second-stage run, both RMSE and RALSD continued to trend downward, with a stronger decline for RALSD. The authors suggest longer training may improve distributional metrics.

A Further Results on the Two-Stage Ablation

Loading 2609.29383v1…