pith. sign in

arxiv: 2606.09959 · v1 · pith:22HLHCHXnew · submitted 2026-06-08 · 💻 cs.LG · cs.AI

Temporal Context Conditioning for Seasonality-Aware Precipitation Nowcasting of High-Intensity Rainfall

Gijs van Nieuwkoop , Siamak Mehrkanoon This is my paper

Pith reviewed 2026-06-27 17:09 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords precipitation nowcastingtemporal conditioningradar dataU-Netcyclical encodingshigh-intensity rainfallseasonal variability
0
0 comments X

The pith

Cyclical encodings of time-of-day and time-of-year improve radar-based nowcasting of high-intensity rainfall when added as conditioning layers to a U-Net.

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

The paper tests whether adding simple temporal context can help deep learning models that nowcast precipitation from recent radar images. It proposes extending a small-attention U-Net with layers that use cyclical encodings of clock time and calendar time to modulate internal features. Experiments on Dutch radar data indicate the added context helps most on rare, intense rainfall events and produces better matches to observed seasonal patterns and intensity distributions. A conductance analysis shows the new layers are actively used despite adding few parameters. These results suggest that lightweight, physically motivated time signals can make nowcasts more realistic without large increases in model size.

Core claim

The Time-Aware SmaAt-UNet (TA-SmaAt-UNet) augments the base SmaAt-UNet with temporal conditioning layers that receive cyclical encodings of time-of-day and time-of-year; when trained on KNMI radar precipitation data the conditioned model yields higher skill on high-intensity events, better seasonal variability, and improved rainfall-intensity histograms compared with the unconditioned baseline.

What carries the argument

Temporal conditioning layers that modulate intermediate feature representations inside the U-Net using cyclical encodings of time-of-day and time-of-year.

If this is right

  • Temporal conditioning yields the largest gains precisely on the rarest, highest-intensity precipitation events.
  • The model produces rainfall-intensity histograms closer to the observed distribution across seasons.
  • Layer conductance analysis confirms the added conditioning layers are utilized during inference.
  • The performance lift occurs while keeping the added parameter count small.

Where Pith is reading between the lines

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

  • The same cyclical-time approach could be tested on nowcasting tasks for other variables such as wind or cloud cover where diurnal and seasonal cycles are strong.
  • Because the encodings are parameter-free and fixed, the method may transfer across different radar networks or geographic regions with minimal retraining.
  • Combining the temporal layers with additional meteorological inputs such as temperature or humidity fields could be examined as a low-cost way to increase context further.

Load-bearing premise

Cyclical encodings of time-of-day and time-of-year supply enough meteorological context to usefully modulate the network's feature representations for rainfall nowcasting.

What would settle it

Retraining the identical SmaAt-UNet architecture on the same KNMI radar sequences but without the temporal conditioning layers and finding no measurable drop in critical success index or intensity distribution match for events above the 95th percentile of rainfall rates.

Figures

Figures reproduced from arXiv: 2606.09959 by Gijs van Nieuwkoop, Siamak Mehrkanoon.

Figure 1
Figure 1. Figure 1: A schematic overview of the proposed TA-SmaAt-UNet model architecture. representation, these temporal quantities are first converted into angular variables, 𝜃d = 2𝜋𝜏d , 𝜃y = 2𝜋 𝑑y 𝑁y , (2) after which the temporal feature vector is defined as 𝐭 = [ sin(𝜃d ), cos(𝜃d ), sin(𝜃y ), cos(𝜃y ) ] ∈ ℝ 4 . (3) This encoding avoids artificial discontinuities at daily and annual boundaries and reflects the periodic na… view at source ↗
Figure 2
Figure 2. Figure 2: MSE and CSI of different models as a function of forecast lead time [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: supports this interpretation from the perspec￾tive of predicted precipitation intensities. The figure shows the ratio between predicted and observed precipitation￾frequency counts across intensity bins. Ideally, all bars [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of ground-truth precipitation and model predictions at selected forecast lead times. Finally, the qualitative comparison in [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Relative module-level conductance of TA-SmaAt￾UNet across forecast lead times for predictions above the 0.5 mm/h threshold. and several temporal baselines sampled from the training set, after which absolute conductance values were averaged. Individual layers were then grouped into four module￾level categories: encoder blocks, CBAM attention modules, temporal conditioning layers (including their associated … view at source ↗
read the original abstract

Precipitation nowcasting is increasingly being approached with deep learning models that learn directly from recent radar observations. Although such models can efficiently capture short-term precipitation motion, they often lack broader contextual information about the meteorological conditions under which rainfall develops. This paper investigates whether lightweight temporal context can improve radar-based nowcasting, particularly for high-intensity rainfall. We propose the Time-Aware Small-Attention U-Net (TA-SmaAt-UNet), which extends the core SmaAt-UNet model with temporal conditioning layers that use cyclical encodings of time-of-day and time-of-year to modulate intermediate feature representations. Experiments on KNMI radar precipitation data show that temporal conditioning is most beneficial for rare, high-intensity precipitation events, while also improving the representation of seasonal variability and predicted rainfall-intensity distributions. A layer conductance analysis further indicates that the added temporal conditioning layers are actively used by the model despite their small parameter cost. These findings suggest that simple, physically motivated temporal context can improve the realism and reliability of deep learning-based precipitation nowcasts. The implementation of our models and training setup is available on \href{https://github.com/gijsvn/TA-SmaAt-UNet}{GitHub}.

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.

Referee Report

0 major / 3 minor

Summary. The paper proposes the Time-Aware Small-Attention U-Net (TA-SmaAt-UNet), which augments the SmaAt-UNet architecture with temporal conditioning layers that apply cyclical encodings of time-of-day and time-of-year to modulate intermediate features. Experiments on KNMI radar precipitation data indicate that this conditioning improves nowcasting performance especially for rare high-intensity events, better captures seasonal variability and rainfall intensity distributions, and that the added layers are utilized according to a conductance analysis. The implementation is released on GitHub.

Significance. If the reported gains hold under the full experimental protocol, the work shows that a lightweight, physically motivated temporal context can measurably improve deep-learning nowcasting of extremes without substantial parameter cost. The open GitHub repository is a clear strength, enabling direct verification of the conditioning mechanism and data pipeline.

minor comments (3)
  1. [Abstract] Abstract: the statement that temporal conditioning is 'most beneficial' for high-intensity events would be strengthened by including one or two key quantitative metrics (e.g., improvement in critical success index or quantile scores) directly in the abstract.
  2. [Methods] The conductance analysis is mentioned as confirming layer utilization; a brief equation or pseudocode in the methods section would clarify how conductance is computed and normalized.
  3. [Results] Figure captions for rainfall-intensity distribution plots should explicitly state the binning scheme and any smoothing applied so that readers can reproduce the comparison.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive summary of the work and for recommending minor revision. The assessment correctly identifies the core contribution of lightweight cyclical temporal conditioning and its benefits for high-intensity events. No major comments were listed in the report, so we have no specific points requiring rebuttal or revision at this stage.

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper's core contribution is an empirical extension of SmaAt-UNet via added temporal conditioning layers that employ standard cyclical encodings of time-of-day and time-of-year. All load-bearing claims (improved performance on high-intensity events, better seasonal representation, and layer utilization) rest on direct experiments, ablations, and conductance analysis on external KNMI radar data rather than any derivation that reduces to a fitted parameter or self-referential definition. No equations, uniqueness theorems, or self-citations are invoked to force the result by construction; the pipeline is fully data-driven and externally verifiable via the linked GitHub code.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that time-of-day and time-of-year encodings capture relevant context for rainfall intensity; no free parameters or invented entities are described in the abstract.

axioms (1)
  • domain assumption Cyclical encodings of time-of-day and time-of-year capture relevant meteorological conditions under which rainfall develops.
    Invoked when proposing the temporal conditioning layers in the abstract.

pith-pipeline@v0.9.1-grok · 5742 in / 1193 out tokens · 16925 ms · 2026-06-27T17:09:58.396045+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    Zängl, D

    G. Zängl, D. Reinert, P. Rípodas, M. Baldauf, The icon (icosahedral non-hydrostatic) modelling framework of dwd and mpi-m: Descrip- tion of the non-hydrostatic dynamical core, Quarterly Journal of the Royal Meteorological Society 141 (2015) 563–579

    2015

  2. [2]

    D. C. Dowell, C. R. Alexander, E. P. James, S. S. Weygandt, S. G. Benjamin, G. S. Manikin, B. T. Blake, J. M. Brown, J. B. Olson, M. Hu, et al., The high-resolution rapid refresh (hrrr): An hourly updating convection-allowing forecast model. part i: Motivation and system description, Weather and Forecasting 37 (2022) 1371–1395

    2022

  3. [3]

    J. A. Brotzge, D. Berchoff, D. L. Carlis, F. H. Carr, R. H. Carr, J. J. Gerth, B. D. Gross, T. M. Hamill, S. E. Haupt, N. Jacobs, et al., Challenges and opportunities in numerical weather prediction, Bulletin of the American Meteorological Society 104 (2023) E698– E705

    2023

  4. [4]

    Prudden, S

    R. Prudden, S. Adams, D. Kangin, N. Robinson, S. Ravuri, S. Mo- hamed, A. Arribas, A review of radar-based nowcasting of precip- itation and applicable machine learning techniques, arXiv preprint arXiv:2005.04988 (2020)

    work page arXiv 2005

  5. [5]

    X. Shi, Z. Chen, H. Wang, D.-Y. Yeung, W.-K. Wong, W.-c. Woo, Convolutional lstm network: A machine learning approach for pre- cipitation nowcasting, Advances in neural information processing systems 28 (2015)

    2015

  6. [6]

    Woo, Deep learning for precipitation nowcasting: A benchmark and a new model, Advances in neural information processing systems 30 (2017)

    X.Shi,Z.Gao,L.Lausen,H.Wang,D.-Y.Yeung,W.-k.Wong,W.-c. Woo, Deep learning for precipitation nowcasting: A benchmark and a new model, Advances in neural information processing systems 30 (2017)

    2017

  7. [7]

    Mehrkanoon, Deep shared representation learning for weather elements forecasting, Knowledge-Based Systems 179 (2019) 120– 128

    S. Mehrkanoon, Deep shared representation learning for weather elements forecasting, Knowledge-Based Systems 179 (2019) 120– 128. G. van Nieuwkoop, S. Mehrkanoon:Preprint submitted to arXivPage 8 of 9 Temporal Context Conditioning for Seasonality-Aware Precipitation Nowcasting of High-Intensity Rainfall

    2019

  8. [8]

    G.Ayzel,T.Scheffer,M.Heistermann, Rainnetv1.0:aconvolutional neural network for radar-based precipitation nowcasting, Geoscien- tific Model Development 13 (2020) 2631–2644

    2020

  9. [9]

    I. A. Abdellaoui, S. Mehrkanoon, Symbolic regression for scientific discovery: an application to wind speed forecasting, in: IEEE sym- posium series on computational intelligence (SSCI), IEEE, 2021, pp. 01–08

    2021

  10. [10]

    Stańczyk, S

    T. Stańczyk, S. Mehrkanoon, Deep graph convolutional networks for wind speed prediction, in: ESANN 2021 Proceedings-29th European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, 2021, pp. 147–152

    2021

  11. [11]

    K.Trebing,T.Stańczyk,S.Mehrkanoon, SmaAt-UNet:Precipitation nowcasting using a small attention-unet architecture, Pattern Recog- nition Letters 145 (2021) 178–186

    2021

  12. [12]

    Trebing, S

    K. Trebing, S. Mehrkanoon, Wind speed prediction using multidi- mensionalconvolutionalneuralnetworks, in:2020IEEEsymposium series on computational intelligence (SSCI), IEEE, 2020, pp. 713– 720

    2020

  13. [13]

    Y. Yang, S. Mehrkanoon, AA-TransUNet: Attention augmented transunet for nowcasting tasks, in: international joint conference on neural networks (IJCNN), IEEE, 2022, pp. 01–08

    2022

  14. [14]

    Aykas, S

    D. Aykas, S. Mehrkanoon, Multistream graph attention networks for wind speed forecasting, in: IEEE Symposium Series on Computa- tional Intelligence (SSCI), IEEE, 2021, pp. 1–8

    2021

  15. [15]

    Kaparakis, S

    C. Kaparakis, S. Mehrkanoon, WF-UNet: Weather data fusion using 3d-unetforprecipitationnowcasting, ProcediaComputerScience222 (2023) 223–232

    2023

  16. [16]

    J. G. Fernández, S. Mehrkanoon, Broad-UNet: Multi-scale feature learningfornowcastingtasks, NeuralNetworks144(2021)419–427

    2021

  17. [17]

    Zhang, M

    Y. Zhang, M. Long, K. Chen, L. Xing, R. Jin, M. I. Jordan, J. Wang, Skilful nowcasting of extreme precipitation with nowcastnet, Nature 619 (2023) 526–532

    2023

  18. [18]

    J. Yin, C. Meo, A. Roy, Z. B. Cher, M. Lică, Y. Wang, R. Imhoff, R. Uijlenhoet, J. Dauwels, Precipitation nowcasting using physics informed discriminator generative models, in: 2024 32nd European SignalProcessingConference(EUSIPCO),IEEE,2024,pp.967–971

    2024

  19. [19]

    L.Vatamány,S.Mehrkanoon,Graphdual-streamconvolutionalatten- tionfusionforprecipitationnowcasting, EngineeringApplicationsof Artificial Intelligence 141 (2025) 109788

    2025

  20. [20]

    Reulen, J

    E. Reulen, J. Shi, S. Mehrkanoon, GA-SmaAt-GNet: Generative adversarialsmallattentiongnetforextremeprecipitationnowcasting, Knowledge-Based Systems 305 (2024) 112612

    2024

  21. [21]

    Dai, Global precipitation and thunderstorm frequencies

    A. Dai, Global precipitation and thunderstorm frequencies. part i: Seasonal and interannual variations, Journal of climate 14 (2001) 1092–1111

    2001

  22. [22]

    Dai, Global precipitation and thunderstorm frequencies

    A. Dai, Global precipitation and thunderstorm frequencies. part ii: Diurnal variations, Journal of Climate 14 (2001) 1112–1128

    2001

  23. [23]

    Z. Gao, X. Shi, H. Wang, Y. Zhu, Y. B. Wang, M. Li, D.-Y. Yeung, Earthformer: Exploring space-time transformers for earth system forecasting, Advances in Neural Information Processing Systems 35 (2022) 25390–25403

    2022

  24. [24]

    Ravuri, K

    S. Ravuri, K. Lenc, M. Willson, D. Kangin, R. Lam, P. Mirowski, M. Fitzsimons, M. Athanassiadou, S. Kashem, S. Madge, et al., Skilful precipitation nowcasting using deep generative models of radar, Nature 597 (2021) 672–677

    2021

  25. [25]

    C. K. Sønderby, L. Espeholt, J. Heek, M. Dehghani, A. Oliver, T. Salimans, S. Agrawal, J. Hickey, N. Kalchbrenner, Metnet: A neural weather model for precipitation forecasting, arXiv preprint arXiv:2003.12140 (2020)

    work page arXiv 2003

  26. [26]

    Espeholt, S

    L. Espeholt, S. Agrawal, C. Sønderby, M. Kumar, J. Heek, C. Bromberg, C. Gazen, R. Carver, M. Andrychowicz, J. Hickey, et al., Deep learning for twelve hour precipitation forecasts, Nature communications 13 (2022) 5145

    2022

  27. [27]

    D. Li, Y. Liu, C. Chen, Msdm v1. 0: A machine learning model for precipitation nowcasting over eastern china using multisource data, Geoscientific Model Development 14 (2021) 4019–4034

    2021

  28. [28]

    J. Ko, K. Lee, H. Hwang, K. Shin, Deep-learning-based precipi- tation nowcasting with ground weather station data and radar data, in: 2022 IEEE International Conference on data mining Workshops (ICDMW), IEEE, 2022, pp. 1063–1070

    2022

  29. [29]

    Leinonen, U

    J. Leinonen, U. Hamann, I. V. Sideris, U. Germann, Thunderstorm nowcasting with deep learning: A multi-hazard data fusion model, Geophysical Research Letters 50 (2023) e2022GL101626

    2023

  30. [30]

    Perez, F

    E. Perez, F. Strub, H. De Vries, V. Dumoulin, A. Courville, Film: Visual reasoning with a general conditioning layer, in: Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018, pp. 3942–3951

    2018

  31. [31]

    Imhoff, C

    R. Imhoff, C. Brauer, A. Overeem, A. Weerts, R. Uijlenhoet, Spatial and temporal evaluation of radar rainfall nowcasting techniques on 1,533events, WaterResourcesResearch56(2020)e2019WR026723

    2020

  32. [32]

    K.Dhamdhere,M.Sundararajan,Q.Yan,Howimportantisaneuron?, arXiv preprint arXiv:1805.12233 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  33. [33]

    Sundararajan, A

    M. Sundararajan, A. Taly, Q. Yan, Axiomatic attribution for deep networks, in: International conference on machine learning, PMLR, 2017, pp. 3319–3328. G. van Nieuwkoop, S. Mehrkanoon:Preprint submitted to arXivPage 9 of 9

    2017