arxiv: 2605.12762 · v1 · submitted 2026-05-12 · 💻 cs.LG · cs.AI

Recognition: 2 theorem links

· Lean Theorem

Multi-Quantile Regression for Extreme Precipitation Downscaling

Hamed Najafi , Gareth Lagerwall , Jayantha Obeysekera , Jason Liu

Authors on Pith no claims yet

Pith reviewed 2026-05-14 21:05 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords precipitation downscalingmulti-quantile regressionextreme eventssuper-resolution networkspinball lossdata augmentationconvective precipitation

0 comments

The pith

A multi-quantile super-resolution network detects extreme precipitation up to 18 times better than deterministic baselines by using pinball loss on separate quantile heads.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard super-resolution networks for precipitation downscaling under-predict heavy-tail events because intensity-weighted MAE averages real and synthetic labels at the same input, shifting the mean rather than learning the conditional distribution. Q-SRDRN addresses this by training separate output heads for quantiles at 0.50, 0.95, 0.99 and 0.999 under pinball loss, with IncrementBound enforcing monotonicity across channels while keeping each head's gradients independent. The design makes cVAE data augmentation complementary: the median head absorbs synthetic samples without contaminating the upper tails. On Florida data the P999 head detects 1598 of 2111 events above 200 mm/day versus 88 for the baseline, with lower KL divergence and RMSE; similar gains appear in California and Texas test regions.

Core claim

Q-SRDRN is a multi-quantile super-resolution network trained with pinball loss at tau in {0.50, 0.95, 0.99, 0.999}. IncrementBound enforces monotonicity across quantile channels while preserving each channel's gradient identity, and separate per-quantile output heads supply independent filter banks for bulk and tail detection. Under this architecture, data augmentation via cVAE becomes complementary because the median head can absorb synthetic patterns without distorting upper-quantile predictions. The result is large gains in extreme-event detection wherever the large-scale precipitation signal is strong.

What carries the argument

Q-SRDRN with IncrementBound monotonicity constraint and separate per-quantile output heads trained under pinball loss, allowing independent optimization of bulk and tail predictions.

If this is right

Extreme precipitation detection improves dramatically in convective, tropical-cyclone and atmospheric-river regimes without sacrificing bulk skill.
Data augmentation can now be applied selectively to the median quantile while tail quantiles remain stable.
Multi-quantile outputs supply calibrated tail probabilities useful for flood-risk modeling.
The method transfers across regions provided the large-scale driver remains strong, even when augmentation does not.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same separate-head design could improve extreme-value modeling for other gridded variables such as wind speed or temperature extremes.
Testing whether the quantile heads can be merged or pruned after training would reduce inference cost while retaining tail accuracy.
Applying the architecture to global reanalysis datasets could reveal whether the gains persist when large-scale signals are weaker.

Load-bearing premise

The main barrier to extreme-event skill is the loss function averaging real and synthetic targets rather than any shortage of training examples or model capacity.

What would settle it

A controlled experiment in which the same cVAE samples added to a standard single-output network trained with intensity-weighted MAE produce no gain in extreme detection rates, or the multi-quantile version loses its advantage on data where the large-scale precipitation signal is weak.

Figures

Figures reproduced from arXiv: 2605.12762 by Gareth Lagerwall, Hamed Najafi, Jason Liu, Jayantha Obeysekera.

**Figure 1.** Figure 1: Q-SRDRN architecture. The shared backbone (16 residual blocks) feeds four separate Conv2D(1) output heads. IncrementBound enforces monotonicity via cumulative softplus increments (Eq. 2). P50 receives pure pinball loss; P95/P99/P999 receive event-weighted pinball loss (α = 5). and βKL=0.008 keeps the posterior informative. For large output grids the fully-connected decoder becomes a bottleneck (e.g., Calif… view at source ↗

**Figure 2.** Figure 2: Calibration gap closure with cVAE augmentation (wet days, y > 1 mm). Bar length is the relative reduction in calibration gap from cVAE augmentation: gap closure = 1 − (raug − 1)/(rbase − 1), where r = Pr(y > qˆτ )/(1 − τ ) is the empirical-to-nominal exceedance ratio (a value of 1 is perfect calibration). Positive (green) bars mean augmentation moves the head toward perfect calibration; negative bars would… view at source ↗

**Figure 3.** Figure 3: CRPS skill score by observed intensity band. Bars show s = (CRPSbase − CRPSaug)/CRPSbase per band: positive (green) bars mean cVAE augmentation improves the pinball-CRPS proxy; negative (red) bars mean it hurts. Underlying CRPS values (mm/day) are printed beneath each bar so the absolute magnitude is recoverable. On Florida, augmentation produces positive skill in every band ≥10 mm/day, peaking at +17.7% i… view at source ↗

**Figure 4.** Figure 4: Sharpness conditioned on observed intensity. Median predictive spread P99 − P50 (circles) and P999 − P95 (squares), binned by observed precipitation. Spread grows monotonically with intensity on both domains, confirming that the upper quantile heads stretch where the data demands. cVAE augmentation narrows the predictive spread (raises sharpness) on the extreme bands of both domains, and the effect is subs… view at source ↗

**Figure 5.** Figure 5: P50 SEDI across precipitation thresholds. SRDRN (single output) vs. Q-SRDRN P50. On CA, P50 alone surpasses SRDRN at ≥100 mm; on FL the un-augmented P50 trails SRDRN at extreme thresholds—a gap closed by augmentation (§4.5). K Matched-Head Envelope Figure This appendix shows the per-head SEDI envelope of Q-SRDRN (no augmentation) with each quantile head plotted only over its natural climatological range: P… view at source ↗

**Figure 6.** Figure 6: Each quantile head in its natural range (Q-SRDRN, no augmentation). Head assignment follows climatological exceedance rates: P50 (blue, ≤10 mm), P95 (green, 10–20 mm), P99 (orange, 30–60 mm), P999 (red, ≥70 mm). The matched-head envelope (black) exceeds the SRDRN baseline (dashed gray) at every threshold on both domains. On California, the P999 head achieves SEDI ≥ 0.99 across the entire ≥50 mm range. 20 … view at source ↗

read the original abstract

Deep super-resolution networks for precipitation downscaling achieve strong bulk skill yet systematically under-predict the heavy-tail events that drive flood risk. We demonstrate that the primary obstacle is the loss function, not the data: under intensity-weighted MAE, real and synthetic labels at the same input are simply averaged, meaning data augmentation shifts the predicted mean rather than the conditional distribution. We resolve this with Q-SRDRN, a multi-quantile super-resolution network trained with pinball loss at tau in 0.50, 0.95, 0.99, 0.999. Two CNN-specific design choices make this practical: IncrementBound enforces monotonicity while preserving each quantile channel's gradient identity, and separate per-quantile output heads provide independent filter banks for bulk and tail detection. Under this design, data augmentation via cVAE becomes complementary: the median head absorbs synthetic patterns without contaminating upper quantiles. Empirically, on Florida (convective/tropical-cyclone dominated), the un-augmented Q-SRDRN P999 head detects 1,598 of 2,111 events at 200 mm/day versus 88 for the deterministic baseline--an 18x detection-rate gain (4.2% to 75.7%)--with 63% lower KL divergence and 3.9% lower RMSE. Adding cVAE-generated samples lifts the P50 channel from 14 to 1,038 hits at 200 mm/day. On California (atmospheric-river dominated), the architecture reaches near-perfect detection (P999 SEDI >= 0.996 through 300 mm/day). On Texas, the baseline catches only 2 of 10,720 events at 200 mm/day while the P999 head catches 8,776 (81.9%). While the cVAE does not transfer across regions, multi-quantile regression captures extremes wherever the large-scale signal is strong, while augmentation rescues the median where it is not.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The multi-quantile pinball setup with separate heads and IncrementBound produces clear gains in extreme-event detection on the three regions, but the shared backbone leaves the claimed isolation of cVAE augmentation unverified.

read the letter

The main thing to know is that switching to pinball loss across multiple quantiles with dedicated output heads lets the model catch far more heavy precipitation events than a standard deterministic network, and the Florida and Texas hit-rate jumps are large enough to notice. On Florida the P999 head goes from 88 to 1,598 detections at 200 mm/day while also cutting KL divergence and RMSE; Texas shows a similar lift from 2 to 8,776 events. That is the practical payoff the paper delivers for flood-relevant tails. The architecture choices are the part that feels new: training separate heads per quantile level, using IncrementBound to keep the outputs monotonic without killing per-channel gradients, and showing that this lets cVAE augmentation improve the median without dragging the upper tails. The abstract walks through why intensity-weighted MAE averages real and synthetic labels together, so the loss change is the lever they pull. They test across convective, atmospheric-river, and other regimes, which gives the results some breadth. The numbers are reported against real observations and a clear baseline, so the empirical side is at least falsifiable. The soft spot is the shared CNN backbone. Separate heads give independent filter banks at the output, but gradients from the median pinball term on augmented samples still flow back through the common layers and can alter features supplied to the P95/P99/P999 heads. IncrementBound addresses monotonicity per channel but does not block that cross-head coupling in the encoder. The paper treats the isolation as solved by the head design, yet without an ablation on feature contamination or gradient norms the complementarity claim stays partly assumptive. The cVAE also fails to transfer across regions, which caps how much the augmentation piece generalizes. This is for people building or using downscaling models who care about extremes rather than bulk statistics. A reader already working on super-resolution precipitation or quantile regression would pick up usable design patterns and see where the gains appear. It is grounded enough and the effect sizes are big enough that a serious editor should send it to referees rather than desk-reject it; the main open question is whether the full methods section shows the isolation holds under closer inspection.

Referee Report

1 major / 0 minor

Summary. The manuscript introduces Q-SRDRN, a multi-quantile super-resolution network for precipitation downscaling trained with pinball loss at tau levels 0.50, 0.95, 0.99, and 0.999. It claims the primary obstacle to capturing heavy-tail events is the loss function rather than data availability, and that two CNN design choices (IncrementBound for monotonicity and separate per-quantile output heads) allow cVAE data augmentation to improve the median without contaminating upper quantiles. Strong empirical gains are reported on three regional datasets, including an 18x increase in detection of 200 mm/day events on Florida data and near-perfect SEDI scores on California.

Significance. If the central claims hold, the work offers a practical route to better extreme-precipitation tails in downscaled fields, which would directly benefit flood-risk and impact modeling. The reported detection-rate and KL-divergence improvements are large enough to be operationally relevant if the isolation between quantile heads can be confirmed.

major comments (1)

[Architecture description] Architecture description: the claim that separate per-quantile output heads supply independent filter banks and thereby prevent median-head gradients from contaminating tail quantiles is undermined by the shared CNN backbone. Gradients from the pinball loss on the median head (especially when trained on cVAE synthetic samples) flow back through the common convolutional layers and can alter feature maps supplied to the P95/P99/P999 heads; IncrementBound is stated to preserve per-channel gradient identity but does not address this cross-head coupling.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and for identifying a key point about gradient flow in the shared backbone. We address the concern directly below and will revise the manuscript to clarify the architecture's behavior while preserving the empirical claims.

read point-by-point responses

Referee: Architecture description: the claim that separate per-quantile output heads supply independent filter banks and thereby prevent median-head gradients from contaminating tail quantiles is undermined by the shared CNN backbone. Gradients from the pinball loss on the median head (especially when trained on cVAE synthetic samples) flow back through the common convolutional layers and can alter feature maps supplied to the P95/P99/P999 heads; IncrementBound is stated to preserve per-channel gradient identity but does not address this cross-head coupling.

Authors: We agree that the shared convolutional backbone permits gradients from the median head (including those arising from cVAE-augmented samples) to propagate into the common feature maps used by the tail heads. The separate output heads do not eliminate this coupling; they only provide quantile-specific parameters after the backbone. However, the design still allows differential specialization: each head applies its own pinball loss, so the optimization can emphasize tail-sensitive features in the later layers even if earlier features are influenced by the median. Empirically, the reported results show that cVAE augmentation improves P50 detection rates while leaving P99/P999 performance stable or improved (e.g., Florida 200 mm/day detection rises from 75.7 % to higher values without degradation of the upper tails). IncrementBound enforces monotonicity per channel but, as noted, does not block cross-head gradient flow. In the revised manuscript we will (i) explicitly acknowledge the shared-backbone coupling, (ii) rephrase the claim to emphasize that separate heads enable per-quantile specialization rather than complete isolation, and (iii) add a short gradient-flow discussion or ablation if space permits. These changes constitute a partial revision. revision: partial

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper introduces Q-SRDRN as an empirical architecture using pinball loss at explicit quantile levels (0.50, 0.95, 0.99, 0.999) together with two CNN design choices (IncrementBound and separate per-quantile heads). All central claims are supported by direct performance measurements against external deterministic baselines and real observations across three regions, with no mathematical derivation, fitted-parameter renaming, or self-citation chain that reduces the reported gains to the inputs by construction. The complementarity of cVAE augmentation is presented as an empirical outcome of the stated design rather than an identity.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The approach rests on the domain assumption that pinball loss will allow independent learning of conditional quantiles in a CNN super-resolution setting and that separate heads prevent cross-contamination between bulk and tail predictions. One new design entity is introduced to enforce monotonicity.

free parameters (1)

quantile levels (0.50, 0.95, 0.99, 0.999)
Explicitly chosen to cover bulk and extreme tails; selection is not data-fitted but author-specified.

axioms (1)

domain assumption Pinball loss enables accurate conditional quantile estimation in this super-resolution setting
Standard property of pinball loss, invoked to justify training separate quantile heads.

invented entities (1)

IncrementBound no independent evidence
purpose: Enforce monotonicity across quantile channels while preserving each channel's gradient identity
New CNN-specific mechanism introduced to make multi-quantile training practical.

pith-pipeline@v0.9.0 · 5673 in / 1634 out tokens · 63540 ms · 2026-05-14T21:05:01.913654+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We resolve this with Q-SRDRN, a multi-quantile super-resolution network trained with pinball loss at τ∈{0.50,0.95,0.99,0.999}. Two CNN-specific design choices make this practical: IncrementBound enforces monotonicity while preserving each quantile channel’s gradient identity, and separate per-quantile output heads provide independent filter banks.
IndisputableMonolith/Foundation/BranchSelection.lean branch_selection unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Under this design, data augmentation via cVAE becomes complementary: the median head absorbs synthetic patterns without contaminating upper quantiles.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages · 3 internal anchors

[1]

Ba\ no-Medina, R

J. Ba\ no-Medina, R. Manzanas, and J. M. Guti \'e rrez. Configuration and intercomparison of deep learning neural models for statistical downscaling. Geoscientific Model Development, 13(4):2109--2124, 2020. DOI: 10.5194/gmd-13-2109-2020 https://doi.org/10.5194/gmd-13-2109-2020

work page doi:10.5194/gmd-13-2109-2020 2020
[2]

K. Bi, L. Xie, H. Zhang, X. Chen, X. Gu, and Q. Tian. Accurate medium-range global weather forecasting with 3 D neural networks. Nature, 619(7970):533--538, 2023. DOI: 10.1038/s41586-023-06185-3 https://doi.org/10.1038/s41586-023-06185-3

work page doi:10.1038/s41586-023-06185-3 2023
[3]

Brando, J

A. Brando, J. Gimeno, J. A. Rodr\'iguez-Serrano, and J. Vitri\`a. Deep non-crossing quantiles through the partial derivative. In Proceedings of the 25th International Conference on Artificial Intelligence and Statistics (AISTATS), volume 151 of PMLR, pages 7902--7914, 2022. arXiv: 2201.12848 https://arxiv.org/abs/2201.12848

work page arXiv 2022
[4]

J. B. Bremnes. Probabilistic forecasts of precipitation in terms of quantiles using NWP model output. Monthly Weather Review, 132(1):338--347, 2004. DOI: 10.1175/1520-0493(2004)132<0338:PFOPIT>2.0.CO;2 https://doi.org/10.1175/1520-0493(2004)132<0338:PFOPIT>2.0.CO;2

work page doi:10.1175/1520-0493(2004)132 2004
[5]

A. J. Cannon. Quantile regression neural networks: Implementation in R and application to precipitation downscaling. Computers & Geosciences, 37(9):1277--1284, 2011. DOI: 10.1016/j.cageo.2010.07.005 https://doi.org/10.1016/j.cageo.2010.07.005

work page doi:10.1016/j.cageo.2010.07.005 2011
[6]

A. J. Cannon. Non-crossing nonlinear regression quantiles by monotone composite quantile regression neural network, with application to rainfall extremes. Stochastic Environmental Research and Risk Assessment, 32(11):3207--3225, 2018. DOI: 10.1007/s00477-018-1573-6 https://doi.org/10.1007/s00477-018-1573-6

work page doi:10.1007/s00477-018-1573-6 2018
[7]

R. E. Chandler and H. S. Wheater. Analysis of rainfall variability using generalized linear models: a case study from the west of Ireland . Water Resources Research, 38(10):1192, 2002. DOI: 10.1029/2001WR000906 https://doi.org/10.1029/2001WR000906

work page doi:10.1029/2001wr000906 2002
[8]

Daniels and M

H. Daniels and M. Velikova. Monotone and partially monotone neural networks. IEEE Transactions on Neural Networks, 21(6):906--917, 2010. DOI: 10.1109/TNN.2010.2044803 https://doi.org/10.1109/TNN.2010.2044803

work page doi:10.1109/tnn.2010.2044803 2010
[9]

C. A. T. Ferro and D. B. Stephenson. Extremal dependence indices: Improved verification measures for deterministic forecasts of rare binary events. Weather and Forecasting, 26(5):699--713, 2011. DOI: 10.1175/WAF-D-10-05030.1 https://doi.org/10.1175/WAF-D-10-05030.1

work page doi:10.1175/waf-d-10-05030.1 2011
[10]

Gal and Z

Y. Gal and Z. Ghahramani. Dropout as a B ayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33rd International Conference on Machine Learning (ICML), volume 48 of PMLR, pages 1050--1059, 2016

work page 2016
[11]

Gneiting and A

T. Gneiting and A. E. Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102(477):359--378, 2007. DOI: 10.1198/016214506000001437 https://doi.org/10.1198/016214506000001437

work page doi:10.1198/016214506000001437 2007
[12]

Harris, A

L. Harris, A. T. T. McRae, M. Chantry, P. D. Dueben, and T. N. Palmer. A generative deep learning approach to stochastic downscaling of precipitation forecasts. Journal of Advances in Modeling Earth Systems, 14(10):e2022MS003120, 2022. DOI: 10.1029/2022MS003120 https://doi.org/10.1029/2022MS003120

work page doi:10.1029/2022ms003120 2022
[13]

D. P. Kingma and M. Welling. Auto-encoding variational B ayes. In International Conference on Learning Representations (ICLR), 2014. arXiv: 1312.6114 https://arxiv.org/abs/1312.6114

work page internal anchor Pith review Pith/arXiv arXiv 2014
[14]

Koenker and K

R. Koenker and K. F. Hallock. Quantile regression. Journal of Economic Perspectives, 15(4):143--156, 2001. DOI: 10.1257/jep.15.4.143 https://doi.org/10.1257/jep.15.4.143

work page doi:10.1257/jep.15.4.143 2001
[15]

Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles

B. Lakshminarayanan, A. Pritzel, and C. Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in Neural Information Processing Systems (NeurIPS), pages 6402--6413, 2017. arXiv: 1612.01474 https://arxiv.org/abs/1612.01474

work page internal anchor Pith review Pith/arXiv arXiv 2017
[16]

H. Liu, S. Zhu, and L. Mo. A novel daily runoff probability density prediction model based on simplified minimal gated memory--non-crossing quantile regression and kernel density estimation. Water, 15(22):3947, 2023. DOI: 10.3390/w15223947 https://doi.org/10.3390/w15223947

work page doi:10.3390/w15223947 2023
[17]

Mardani, N

M. Mardani, N. Brenowitz, Y. Cohen, J. Pathak, C.-Y. Chen, C.-C. Liu, A. Vahdat, M. A. Nabian, T. Ge, A. Subramaniam, K. Kashinath, J. Kautz, and M. Pritchard. Residual corrective diffusion modeling for km-scale atmospheric downscaling. Communications Earth & Environment, 6:124, 2025. DOI: 10.1038/s43247-025-02042-5 https://doi.org/10.1038/s43247-025-02042-5

work page doi:10.1038/s43247-025-02042-5 2025
[18]

Vandal, E

T. Vandal, E. Kodra, J. Dy, S. Ganguly, R. Nemani, and A. R. Ganguly. Quantifying uncertainty in discrete-continuous and skewed data with B ayesian deep learning. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD), pages 2377--2386, 2018. DOI: 10.1145/3219819.3219996 https://doi.org/10.1145/3219819.3219996

work page doi:10.1145/3219819.3219996 2018
[19]

In: Proceedings of the 40th International Conference on Machine Learning

T. Nguyen, J. Brandstetter, A. Kapoor, J. K. Gupta, and A. Grover. ClimaX : A foundation model for weather and climate. In Proceedings of the 40th International Conference on Machine Learning (ICML), pages 25904--25938, 2023. arXiv: 2301.10343 https://arxiv.org/abs/2301.10343

work page arXiv 2023
[20]

FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators

J. Pathak, S. Subramanian, P. Harrington, S. Raja, A. Chattopadhyay, M. Mardani, T. Kurth, D. Hall, Z. Li, K. Azizzadenesheli, P. Hassanzadeh, K. Kashinath, and A. Anandkumar. FourCastNet : A global data-driven high-resolution weather model using adaptive F ourier neural operators. arXiv preprint, 2022. arXiv: 2202.11214 https://arxiv.org/abs/2202.11214

work page internal anchor Pith review Pith/arXiv arXiv 2022
[21]

Price and S

I. Price and S. Rasp. Increasing the accuracy and resolution of precipitation forecasts using deep generative models. In Proceedings of the 25th International Conference on Artificial Intelligence and Statistics (AISTATS), volume 151 of PMLR, pages 10555--10571, 2022

work page 2022
[22]

Nature637, 84–90 (2025) https://doi.org/10.1038/s41586-024-08252-9

I. Price, A. Sanchez-Gonzalez, F. Alet, T. R. Andersson, A. El-Kadi, D. Masters, T. Ewalds, J. Stott, S. Mohamed, P. Battaglia, R. Lam, and M. Willson. Probabilistic weather forecasting with machine learning. Nature, 637(8044):84--90, 2025. DOI: 10.1038/s41586-024-08252-9 https://doi.org/10.1038/s41586-024-08252-9

work page doi:10.1038/s41586-024-08252-9 2025
[23]

Ravuri, K

S. Ravuri, K. Lenc, M. Willson, D. Kangin, R. Lam, P. Mirowski, M. Fitzsimons, M. Athanassiadou, S. Kashem, S. Madge, R. Prudden, A. Mandhane, A. Clark, A. Brock, K. Simonyan, R. Hadsell, N. Robinson, E. Clancy, A. Arribas, and S. Mohamed. Skilful precipitation nowcasting using deep generative models of radar. Nature, 597(7878):672--677, 2021. DOI: 10.103...

work page doi:10.1038/s41586-021-03854-z 2021
[24]

Y. Sha, D. J. Gagne II, G. West, and R. Stull. Deep-learning-based gridded downscaling of surface meteorological variables in complex terrain. Part I : Daily maximum and minimum 2-m temperature. Journal of Applied Meteorology and Climatology, 59(12):2057--2073, 2020. DOI: 10.1175/JAMC-D-20-0057.1 https://doi.org/10.1175/JAMC-D-20-0057.1

work page doi:10.1175/jamc-d-20-0057.1 2057
[25]

K. Sohn, X. Yan, and H. Lee. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems (NeurIPS), pages 3483--3491, 2015

work page 2015
[26]

Vandal, E

T. Vandal, E. Kodra, S. Ganguly, A. Michaelis, R. Nemani, and A. R. Ganguly. DeepSD : Generating high resolution climate change projections through single image super-resolution. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), pages 1663--1672, 2017. DOI: 10.1145/3097983.3098004 https://doi.org/...

work page doi:10.1145/3097983.3098004 2017
[27]

A. W. Wood, L. R. Leung, V. Sridhar, and D. P. Lettenmaier. Hydrologic implications of dynamical and statistical approaches to downscaling climate model outputs. Climatic Change, 62(1--3):189--216, 2004. DOI: 10.1023/B:CLIM.0000013685.99609.9e https://doi.org/10.1023/B:CLIM.0000013685.99609.9e

work page doi:10.1023/b:clim.0000013685.99609.9e 2004
[28]

F. Wang, D. Tian, L. Lowe, L. Kalin, and J. Lehrter. Deep learning for daily precipitation and temperature downscaling. Water Resources Research, 57(4):e2020WR029308, 2021. DOI: 10.1029/2020WR029308 https://doi.org/10.1029/2020WR029308

work page doi:10.1029/2020wr029308 2021
[29]

Wehenkel and G

A. Wehenkel and G. Louppe. Unconstrained monotonic neural networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 1543--1553, 2019. arXiv: 1908.05164 https://arxiv.org/abs/1908.05164

work page arXiv 2019
[30]

Zorita and H

E. Zorita and H. von Storch. The analog method as a simple statistical downscaling technique: comparison with more complicated methods. Journal of Climate, 12(8):2474--2489, 1999. DOI: 10.1175/1520-0442(1999)012<2474:TAMAAS>2.0.CO;2 https://doi.org/10.1175/1520-0442(1999)012<2474:TAMAAS>2.0.CO;2

work page doi:10.1175/1520-0442(1999)012 1999