Physical Scales Matter: The Role of Receptive Fields and Advection in Satellite-Based Thunderstorm Nowcasting with Convolutional Neural Networks
Christoph Metzl, \aff a Kianusch Vahid Yousefnia, \aff a Richard Müller, \aff b Virginia Poli, \aff c \aff d Miria Celano, \aff c Tobias Bölle, \aff a · 2025
Read Physical Scales Matter: The Role of Receptive Fields and Advection in Satellite-Based Thunderstorm Nowcasting with Convolutional Neural Networks by Christoph Metzl, \aff a Kianusch Vahid Yousefnia, \aff a Richard Müller, \aff b Virginia Poli, \aff c \aff d Miria Celano, \aff c Tobias Bölle, \aff a (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-09994.txt.
Introduction
Christoph Metzl and co-authors present a 2025 paper on storm forecasts from satellite images. Short-range storm prediction is shifting from physics-based drift methods to pure machine learning. Adding drift back in helped radar rain forecasts, but whether it helps satellite storm forecasts was untested. The team offers a scale argument for when the gain should appear.
Two networks were trained to mark lightning zones. The plain one saw short image series alone. The drift-aware one also got a drifted forecast of each input at the target time. Average scores barely differed. Yet split by forecast length and wind speed, the drift-aware model pulled ahead after two hours, just as the scale argument predicted. Physics scales should shape model design, the authors argue.
Catalog record
| CREATOR | Christoph Metzl, \aff a Kianusch Vahid Yousefnia, \aff a Richard Müller, \aff b Virginia Poli, \aff c \aff d Miria Celano, \aff c Tobias Bölle, \aff a |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 65 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-09994
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Physical Scales Matter: The Role of Receptive Fields and Advection in Satellite-Based Thunderstorm Nowcasting with Convolutional Neural Networks Christoph Metzl, \aff a Kianusch Vahid Yousefnia, \aff a Richard Müller, \aff b Virginia Poli, \aff c \aff d Miria Celano, \aff c Tobias Bölle, \aff a Abstract The focus of nowcasting development is transitioning from physically motivated advection methods to purely data-driven Machine Learning (ML) approaches. Nevertheless, recent work indicates that incorporating advection into the ML value chain has improved skill for radar-based precipitation nowcasts. However, the generality of this approach and the underlying causes remain unexplored. This study investigates the generality by probing the approach on satellite-based thunderstorm nowcasts for the first time. Resorting to a scale argument, we then put forth an explanation when and why skill improvements can be expected. In essence, advection guarantees that thunderstorm patterns relevant for nowcasting are contained in the receptive field at long forecast times. To test our hypotheses, we train ResU-Nets solving segmentation tasks with lightning observations as ground truth. The input of the Baseline Neural Network (BNN) are short time series of multispectral satellite imagery and lightning observations, whereas the Advection-Informed Neural Network (AINN) additionally receives the Lagrangian persistence nowcast of all input channels at the desired forecast time. Overall, we find only a minor skill improvement of the AINN over the BNN when considering fully averaged scores. However, assessing skill conditioned on forecast time and advection speed, we demonstrate that our scale argument correctly predicts the onset of skill improvement of the AINN over the BNN after 2 h 2\text{\,}\mathrm{h} forecast time. We confirm that, generally, advection becomes gradually more important with longer forecast times and higher advection speeds. Our work accentuates the importance of considering and incorporating the underlying physical scales when designing ML-based forecasting models. \statement This research tests the effectiveness of combining physically motivated algorithms with artificial intelligence for short-term thunderstorm forecasts. Our neural network uses geostationary satellite imagery and lightning observations together with their advection forecasts as inputs for predicting the probability of lightning occurrence. Additionally, we propose a simple scale-based explanation for when and why incorporating physically motivated algorithms benefit convolutional neural networks, and we test this hypothesis. We find that our scale argument can roughly predict the onset of usefulness. The application of the scale argument is quite general and can aid in the development of improved forecasting models in multiple domains. 1 Introduction Extreme weather events related to thunderstorm activity such as lightning, flash floods, hail, and strong winds, are responsible for severe economic loss ( Hoeppe, 2016 ) and can be a threat to human life ( Terti et al., 2017 ) . Increases in severity and frequency of thunderstorms driven by climate change are expected ( Diffenbaugh et al., 2013 ; Rädler et al., 2019 ; Raupach et al., 2021 ) . This necessitates timely and accurate very short-term forecasts also known as nowcasts ( Bojinski et al., 2023 ) . Physically motivated nowcasting methods rely on advection and empirical handcrafted rules derived from experience ( Wang et al., 2017 ) and are at present dominant in operational use ( James et al., 2018 ; Pulkkinen et al., 2019 ; Pulkkinen et al., 2020 ; Müller et al., 2022 ) . Advection requires an estimate of the (large-scale) ambient wind and is typically implemented by solving the optical-flow problem (equation ( 4 ) for s_{c}=0 ) for a sequence of remote-sensing images ( Lucas and Kanade, 1981 ; Bowler et al., 2006 ; Santek et al., 2019 ; Apke et al., 2022 ) . The data-driven nature of nowcasting implies ML methods to be a promising alternative. Indeed, the main focus in the development of state-of-the-art nowcasting techniques relies on deep learning specifically ( Ravuri et al., 2021 ; Zhang et al., 2023 ) . However, it is currently unsettled how to make the best use of ML in this context. Hence, prior work mainly focuses on optimizing the ML workflow as seen in the strong heterogeneity across all relevant parts of the value chain: Considered hazards include precipitation ( Zhao et al., 2024 ) , lightning ( Geng et al., 2021 ) and hail ( Leinonen et al., 2023b ) . Ground truth candidates are direct ( Wang et al., 2023 ) , thresholded ( Ortland et al., 2023 ) observations or human labels ( Cintineo et al., 2020 ) . ML tasks scan classification ( Vahid Yousefnia et al., 2024b ) , segmentation ( Lagerquist et al., 2021 ) , supervised ( Guo et al., 2023 ) and generative regression ( Leinonen et al., 2023a ) . Data sources encompass various combinations of radar ( Trebing et al., 2021 ) , satellite ( Brodehl et al., 2022 ) and lightning ( Zhou et al., 2020 ) observations as well as weather stations ( Andrychowicz et al., 2023 ) , numerical weather prediction ( Vahid Yousefnia et al., 2024a ) or digital elevation models ( Leinonen et al., 2022a ) . Strategies to address rare events or class imbalances manifest either in different under-sampling strategies in the dataset composition ( Ayzel et al., 2020 ) ), in class weighting during model training ( Bi et al., 2023 ) , or in other loss function adaptations ( Yang and Yuan, 2023 ) . Furthermore, model architectures leverage building blocks of feedforward ( Vahid Yousefnia et al., 2025 ) , recurrent ( Shi et al., 2015 ) , convolutional ( Han et al., 2022 ) , graph neural networks ( Farahmand et al., 2023 ) and Transformers ( Bai et al., 2022 ) . Recent works indicate that including physics-informed domain knowledge in the form of advection directly into the ML value chain is a promising complementary route to improve nowcasting ( Zhang et al., 2023 ; Ritvanen et al., 2023 ; Ha and Lee, 2023 ; Pavlík et al., 2024 ) . All prior work in that direction focuses exclusively on radar-based precipitation nowcasting, but differs in the implementation details. Zhang et al. (2023) design a separate evolution network tasked with learning the advection, whose outputs are then conditioned upon and refined by a generative adversarial network (GAN). In order to obtain a fully end-to-end optimizable model, they do not utilize traditional optical flow algorithms. Ha and Lee (2023) also build upon the advection but utilize multiple traditional optical flow algorithms by combining them with a regression model, which in turn serves as an additional input for a U-Net. Ritvanen et al. (2023) pursue a separate approach of employing a traditional optical flow algorithm to detach the growth and decay of rainfall from the advection. They achieve this by training a U-Net to learn the approximated time derivative of the precipitation field in the Lagrangian coordinates. Pavlík et al. (2024) extend the approach of Ritvanen et al. (2023) . They keep the U-Net for learning the growth and decay but implement the transformation to Lagrangian coordinates in an end-to-end trainable fashion similar to Zhang et al. (2023) . The first aim of this work is to test the generality of the concept of combining advection with ML by shifting the hazard type from precipitation to lightning, the ML task from regression