Teufel et al.
Geoscience Letters (2023) 10:19
[Link] Geoscience Letters
RESEARCH LETTER Open Access
Physics‑informed deep learning framework
to model intense precipitation events at super
resolution
B. Teufel1* , F. Carmo1, L. Sushama1, L. Sun1, M. N. Khaliq2, S. Bélair3, A. Shamseldin4, D. Nagesh Kumar5 and
J. Vaze6
Abstract
Physical modeling of precipitation at fine (sub-kilometer) spatial scales is computationally very expensive. This study
develops a highly efficient framework for this task by coupling deep learning (DL) and physical modeling. This frame-
work is developed and tested using regional climate simulations performed over a domain covering Montreal and
adjoining regions, for the summers of 2015–2020, at 2.5 km and 250 m resolutions. The DL framework uses a recurrent
approach and considers atmospheric physical processes, such as advection, to generate high-resolution information
from low-resolution data, which enables it to recreate fine details and produce temporally consistent fields. The DL
framework generates realistic high-resolution precipitation estimates, including intense short-duration precipitation
events, which allows it to be applied in engineering problems, such as evaluating the climate resiliency of urban
storm drainage systems. The results portray the value of the proposed DL framework, which can be extended to other
resolutions, periods, and regions.
Keywords Deep learning, Regional climate model, Convection permitting model, Intense precipitation, Engineering
scale
Introduction storm water drainage. In a warming climate, the intensity
Intense short-duration precipitation events can lead to and frequency of extreme precipitation are expected to
catastrophic flash flooding in urban regions and their increase due to the increasing water-holding capacity of
intensity/duration/frequency characteristics are thus the atmosphere (e.g., IPCC 2013), which is increasingly
widely used in the design of engineering systems, such as being accounted for by practitioners during the design
of civil infrastructure. Understanding projected changes
to short-duration precipitation extremes is particularly
*Correspondence:
B. Teufel
important, given that rapid urbanization is occurring,
[Link]@[Link] leading to over two-thirds of the global population being
1
Department of Civil Engineering, Trottier Institute for Sustainability projected to live in urban regions by 2050 (UN 2018).
in Engineering and Design, McGill University, Montreal, Canada
2
Ocean, Coastal and River Engineering (OCRE) Research Centre, National
Urban regions influence regional weather and climate,
Research Council Canada, Ottawa, Canada as they are replete with anthropogenic heat and aero-
3
Meteorological Research Division, Science and Technology Branch, sol sources, store little water and obstruct atmospheric
Environment and Climate Change Canada, Dorval, Canada
4
Department of Civil and Environmental Engineering, University
motion (e.g., Oke 1982; Huszar et al. 2014; Daniel et al.
of Auckland, Auckland, New Zealand 2019). The urban heat island (UHI), characterized by
5
Department of Civil Engineering, Indian Institute of Science, Bangalore, higher temperatures in urban regions than surrounding
India
6
CSIRO Land and Water, Canberra, Australia
areas, may contribute to strengthening convection which
© Crown 2023. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits
use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original
author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third
party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the mate-
rial. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation
or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit
[Link]
Teufel et al. Geoscience Letters (2023) 10:19 Page 2 of 9
can lead to enhanced precipitation extremes (e.g., Shep- 2009; Trinh et al. 2014; Vandal et al. 2017; Xie et al. 2018;
herd and Burian 2003; Mölders and Olson 2004). Stengel et al. 2020). The CNN-based SR approach is data-
Regional climate models (RCMs) are useful tools to driven and does not require solving complex analytical
study projected changes to climate, which are being formulations, which dramatically decreases the compu-
increasingly used to inform effective adaptation meas- tational cost of generating high-resolution data once the
ures to cope with the impacts of global warming in many parameters of CNNs are properly trained and validated.
fields. However, short-duration precipitation extremes This study focuses on the development, validation
are generally associated with small-scale processes, such and application of a novel framework combining the
as deep convection, which are not resolved at the scales physically based regional climate model GEM (Global
of most existing RCM simulations and are thus approxi- Environmental Multiscale) outputs with deep learning
mated through various parameterizations. The recent techniques to efficiently generate precipitation informa-
studies have shown that high resolution (i.e., convection tion at applied engineering-oriented fine spatial scales.
permitting) RCM simulations can realistically capture Such fine resolution information is extremely useful for
short-duration precipitation extremes due to better rep- evaluating climate resiliency of urban infrastructure sys-
resentation of mesoscale dynamics, cloud microphysics, tems to flash flooding in urban centers such as Montreal,
surface heterogeneity and orographic effects (Prein et al. where flash flooding caused by high-intensity rainfall is
2015; Kendon et al. 2017; Diro and Sushama 2019; Teufel a recurrent problem. Montreal is the second largest city
and Sushama 2022). of Canada and is the major economic hub in eastern
Development of effective adaptation and mitigation Canada. Weather related disruptions can trigger cascad-
strategies requires information of intense precipitation ing impacts on various interconnected urban systems in
changes at engineering/super scales (i.e., < 1 km spatial Montreal. Given such a high impact of short-duration
resolution). The previous studies have shown that cli- precipitation extremes in urban regions, the objective of
mate simulations with advanced representation of urban this study is to develop an efficient and physically consist-
regions in RCMs are able to adequately capture urban− ent emulator for intense precipitation during the summer
climate feedbacks (e.g., Teufel et al. 2021). However, season for Montreal, at the sub-hourly time scales associ-
despite significant developments in computing technol- ated with deep convection, to support studies on climate
ogy, parallel programming architectures and improved resilience of various urban systems. In addition, the goal
representation of physical processes in climate models, is to keep the development as generic as possible so that
high computational cost continues to be a major bar- the developed framework can be extended to other sea-
rier in undertaking climate simulations at engineering sons, periods and regions of the world.
scales for sufficiently longer periods. For example, a cli- The remainder of this manuscript is organized as fol-
mate simulation at 250 m requires about 1000 times the lows: “Methodology” section describes the machine
computing resources required at 2.5 km over the same learning framework, the physical climate model, and the
domain. In addition, ensembles of climate simulations experiments performed. “Results” section presents the
are generally required to quantify uncertainty in cli- outputs obtained from the machine learning model and
mate projections and that further amplifies the required compares them to the climate model outputs. Finally,
resources. The recent advances in machine learning (Bre- “Discussion and conclusions” section provides discus-
nowitz and Bretherton 2018; Reichstein et al. 2019) and sion, conclusions and future avenues of research. Main
its applications in various fields (e.g., Chung and Shin highlights of the study are also discussed in this section.
2018; Ding et al. 2020; Pradhan et al. 2020; Stengel et al.
2020; Van et al. 2020; Ray and Chattopadhyay 2021; Bar- Methodology
rera-Animas et al. 2022; Girihagama et al. 2022) provide Deep learning framework
an opportunity for developing hybrid approaches, com- Given that precipitation evolves in both space and time,
bining physical understanding of atmospheric processes the objective of generating high-resolution (HR) precipi-
with machine learning architectures, to overcome this tation from low-resolution (LR) precipitation is func-
obstacle and advance studies on climate−urban infra- tionally similar to the problem of video super-resolution
structure interactions (Wu et al. 2021) and informing (VSR). VSR differs from single image SR in that temporal
design methodologies. In the recent years, deep learning- relationships are exploited by combining the information
based image super-resolution (SR) models built using from multiple LR frames to reach better quality results.
convolutional neural networks (CNNs) have been devel- Many existing VSR methods approach the problem by
oped (e.g., Wang et al. 2015; Dong et al. 2016; Lai et al. combining a batch of LR frames to estimate a single HR
2017; Zhang et al. 2018) and applied to produce high- frame, effectively dividing the task of VSR into a large
resolution physical fields in various domains (Li et al. number of separate multiframe SR subtasks (Caballero
Teufel et al. Geoscience Letters (2023) 10:19 Page 3 of 9
et al. 2017; Liu et al. 2017; Makansi et al. 2017; Tao et al. videos, significantly outperforming other state of the art
2017). However, generating each output frame separately methods.
reduces the method’s ability to produce temporally con- In this study, two physical considerations are taken
sistent HR frames, often resulting in artifacts. into account to adapt the FRVSR approach to the task of
The deep learning (DL) framework used in this study generating SR precipitation. First, in VSR the displace-
is inspired by the one proposed by Sajjadi et al. (2018), ments from one frame to the next (called optical flow)
named frame-recurrent video super-resolution (FRVSR; are typically estimated from the frames themselves—Saj-
Fig. 1a). This recurrent approach passes the previ- jadi et al. (2018) used an encoder−decoder CNN for this
ously estimated HR frame as an input for the follow- task. The same approach could be chosen for SR precipi-
ing iteration. Information from past frames can thus tation, but instead, the physical fact that precipitation is
be propagated to later frames which helps the model to advected by wind is used to derive optical flow, which has
recreate fine details and produce temporally consistent the advantage that flow estimates are available at all times
Fig. 1 a The architecture of the FRVSR framework by Sajjadi et al. (2018). Blue boxes are used for LR input data, yellow boxes for field operators,
green boxes for HR data and red boxes for CNNs. Loss terms are shown in orange. b Proposed deep learning framework for precipitation. c GEM
computational domain at 2.5 km (brown; every 5th grid cell shown) and 250 m (blue; every 20th grid cell shown). The outer thick lines represent the
model domain, while the inner thick lines represent the model free domain. Background colors represent orography (m asl). d Urban fraction in the
250 m domain. Water bodies are shown in dark blue
Teufel et al. Geoscience Letters (2023) 10:19 Page 4 of 9
and locations, given that wind is a continuous field. This are shown in Fig. 1c, with the smaller 250 m resolution
approach also avoids having to train an additional CNN domain covering the city of Montreal. The urban cover-
and needing to merge two loss terms. Second, the CNNs age for Montreal and surroundings is shown in Fig. 1d.
in Sajjadi et al. (2018) operate in LR space, enabling faster GEM is used for numerical weather prediction at Envi-
training, but imposing some constraints on the choice of ronment and Climate Change Canada (ECCC). The land
the upscale factor (4 × in their case). This is undesirable part of the model is represented using the Canadian
in SR precipitation, where the upscale factor may vary Land Surface Scheme–CLASS (Verseghy 2011), while
depending on the specific resolutions of the HR and LR the urban regions are represented by the Town Energy
climate simulations. In addition, applying transposed Balance (TEB; Masson 2000) model. Condensation pro-
convolutions or similar methods to generate HR outputs cesses are computed by a double-moment microphys-
can result in undesirable gradients and artifacts—for ics scheme (Milbrandt and Yau 2005). More details on
these reasons, the CNN in the proposed SR precipitation the parameterizations used can be found in Diro and
framework is set to operate in HR. Sushama (2019). The urban climate simulation at high
The estimates of FRVSR tend to improve with time, resolution is driven at the lateral boundaries by the low
i.e., for the first frame of a scene, its performance can- resolution GEM simulation, which is in turn driven by
not exceed that of a single-frame image reconstruc- ERA5 reanalysis data (Hersbach et al. 2020) from the
tion method. For subsequent frames, the framework is European Centre for Medium-Range Weather Forecasts.
able to leverage past information, thus producing better GEM outputs at 1-min intervals from both simulations
estimates. One implication of this is that if movement constitute the input data for the DL framework, and
between consecutive frames is too fast, each frame will match the timestep of the LR simulation. The HR simu-
appear relatively independent from the previous ones, lation uses a 10 s timestep, but the highest feasible tem-
and a significant fraction of the new frame will not have poral resolution is constrained by the coarser timestep of
been previously seen by the DL model, which would the LR simulation (60 s).
degrade its performance. To avoid this, displacements
between frames should be small—in the case of SR pre- Deep learning experiments
cipitation, this can be achieved by using the highest feasi- It is fundamental to note that the HR GEM simulation
ble temporal resolution, as constrained by the timestep of is driven by the LR GEM simulation only at the lateral
the climate model. boundaries, which means that the HR simulation has
Figure 1b shows the DL framework used in this study. considerable freedom to develop its own precipitation
The inputs consist of LR precipitation fields and LR evolution, which often differs from the LR simulation in
storm motion estimates (derived from cloud-level winds). precipitation coverage and intensity. Although expected,
LR information first needs to be upscaled to HR, which is these differences make it unadvisable to train the DL
accomplished using bilinear interpolation on precipita- framework using the HR and LR simulation outputs
tion and each storm motion component (u, v). The HR directly. Instead, a two step process is developed: first,
storm motion is then used to advance the previous HR the DL framework is trained using the HR data and LR
precipitation estimate in time, which is then used along- data generated from it (this is the method by which VSR
side the upscaled LR precipitation as inputs to SRNet, frameworks are usually trained); and second, the trained
which outputs the new best estimate of HR precipita- framework is applied to the LR simulation outputs to
tion. The above procedure is then repeated for each sub- generate HR data.
sequent timestep. SRNet is a 5-layer CNN in which the For the first step, it is desirable to generate LR data in
convolutional kernel of each layer is 3 × 3 and each layer which precipitation follows the spatiotemporal evolution
outputs 16 feature maps. The rectified linear unit (ReLU) of the HR data as closely as possible, given that during
function is used as the activation function for all convo- training the DL framework will strive to minimize errors
lutional layers. The loss function is defined as the mean in both space and time in the reconstruction of HR pre-
squared error (MSE) of the reconstruction (for all grid cipitation. This generated LR data (LRG) is obtained by
cells and all timesteps). simply averaging the HR fields over each LR grid cell.
Here, the data for the summers of 2015–2018 is used as
Climate model and simulations the training dataset, and the data for 2019 is used as the
The limited area version of the physically based model validation dataset. Finally, the data for 2020 are used as
GEM (Côté et al. 1998; Girard et al. 2014) is used to the test dataset.
perform simulations for the summers (June–August) For the second step, the unseen LR simulations out-
of 2015–2020, at 2.5 km (low) and 250 m (high) resolu- puts for the entire 2015–2020 period are used as inputs
tion. The experimental domains for the two resolutions to the trained DL framework. The HR outputs from the
Teufel et al. Geoscience Letters (2023) 10:19 Page 5 of 9
framework will closely follow the spatiotemporal evo- are performed both for select intense precipitation
lution of the LR data, which differs from the HR truth. events, and in terms of overall statistics for the testing
Given this, and the focus on high-intensity precipitation period.
events, the performance of the framework needs to be Figure 2a shows the precipitation intensities for
assessed in terms of statistical similarity, such as the abil- 1-min snapshots during selected heavy precipita-
ity to reproduce the precipitation frequency–intensity tion events. It can be seen that the DL framework
relationships, often used for the design of urban drainage performs very well compared to the HR truth, being
networks. able to reproduce many details not visible in the LRG
input for all of the events. For instance, the HR truth
Results for event 1 shows the presence of many cores (local
Reconstruction of HR information from LRG data maxima) of heavy precipitation. The DL framework is
The performance of the DL framework is assessed by able to realistically capture most of those cores, being
comparing its estimates to the HR truth, for unseen data only slightly smoother than the HR truth. The same is
in the testing period (i.e., summer of 2020). Comparisons true for events 3 and 4, although for the latter, the finest
Fig. 2 a Precipitation fields (mm/h) for select high-intensity precipitation events, from L RG (top), DL model (middle) and HR truth (bottom). b
Frequency–intensity relationships for the three datasets in a, over the entire HR domain. c As b, but for Montreal’s downtown core
Teufel et al. Geoscience Letters (2023) 10:19 Page 6 of 9
structures (only few HR gridcells wide) are not fully Generation of HR information from actual LR data
reproduced. The DL framework also performs very well Given that the developed DL framework performs well
at delineating the regions with precipitation intensities at reconstructing the HR truth from the LRG data (see
above 1 mm/h and it is also able to emulate the spatial “Reconstruction of HR information from LRG data” sec-
structure of precipitation cores, in which high-inten- tion), the next step is to apply the framework to the actual
sity precipitation is surrounded by lesser precipitation LR data and assess the realism of the generated output
intensities, a feature not seen in LRG, where a grid cell in terms of its similarity to the HR truth. Given that the
with high-intensity precipitation can be adjacent to one fields cannot be directly compared at the grid cell level
with zero or negligible precipitation. due to the differences between LR and HR GEM, meas-
To further assess its performance, the ability of the ures of statistical similarity are used.
DL framework to reproduce the precipitation inten- GEM simulations at LR exhibit a significantly greater
sity−frequency relationship is evaluated. In Fig. 2b, degree of spatial autocorrelation than the L RG data cre-
it can be seen that the frequency of very heavy pre- ated from HR fields (Fig. 3a). This is not surprising, con-
cipitation (above 20 mm/h) is underestimated when sidering that the actual resolution of a climate model is
averaging the HR data over a LR grid. For the heavi- around 4–6 times coarser than its grid spacing, which
est intensities (100 mm/h and above), the frequency means that the L RG data are expected to have a greater
is underestimated by up to 50%. The DL framework is amount of spatial detail (originally resolved at HR) than
able to well reconstruct the frequency of heavy precipi- the corresponding LR data.
tation, reversing most of the underestimation present Given that the DL framework is trained using the L RG
in the LRG input, and being close to the HR truth over data, it is desirable to transform the actual LR data so
the entire range of precipitation intensities. Since the that it better resembles the L RG data. The transforma-
performance of the framework for specific locations tion applied reduces spatial autocorrelation (Fig. 3a) and
within the domain is also of interest, Fig. 2c shows the is functionally similar to a sharpening filter. The transfor-
intensity−frequency curves for Montreal’s downtown mation is performed by convolution with a 3 × 3 kernel,
core. Compared with Fig. 2b, the underestimation of in which the center element has a value of 1.88, the four
the heaviest precipitation rates is more pronounced in corners are zero, and the other four elements are − 0.22
LRG, reaching close to 80%. The DL framework is able (the sum of all elements equals one as to not introduce
to correct a significant fraction of the underestimation, bias). These values are chosen with the goal of improv-
but not its entirety. One potential explanation would be ing the frequency−intensity relationship of the LR data,
higher frequency of precipitation events with very fine which is much closer to LRG after the transformation
spatial structures over the downtown core, which are (Fig. 3b).
not fully reproduced by the DL framework (as previ- Applying the trained DL model to the transformed
ously mentioned). LR data produces HR estimates that closely follow the
Fig. 3 a Median spatial autocorrelations in LRG, LR (GEM) and LR (transformed), and their differences with respect to LRG. b Frequency–intensity
relationships for the three datasets in a
Teufel et al. Geoscience Letters (2023) 10:19 Page 7 of 9
spatiotemporal evolution of precipitation in LR GEM, Discussion and conclusions
but with the statistical characteristics of HR GEM. Fig- In this study, a physics-informed deep learning frame-
ure 4a shows that LR GEM significantly underestimates work that enhances the resolution of precipitation from
the frequency of occurrence of torrential precipita- 2.5 km to 250 m (i.e., by a factor of 10) is developed and
tion (above 50 mm/h), which is greatly improved (i.e., tested for Montreal, which is the second largest city in
much closer to the HR truth) after applying the DL Canada, using outputs from a physically based regional
framework. It is worth noting that the locations of high climate model. It is noted that the deep learning frame-
values in the DL outputs resemble those of LR GEM, work is able to recreate fine details and produce tempo-
which are not the same as in HR GEM. Given the rela- rally consistent precipitation fields by taking into account
tive rarity of such heavy precipitation, the differences physical atmospheric processes, such as the advection of
in the location of the maxima are a consequence to the precipitation by wind, which is the novelty of this study.
small sample size (6 years) and not indicative of long- The results show that the deep learning model is capable
term means. In terms of the precipitation intensity− of capturing many of the fundamental characteristics of
frequency relationship (Fig. 4b), the DL framework intense short-duration precipitation events at fine spa-
produces results that are very close to the HR truth tial scales, such as their coverage, intensity and spatial
over the entire intensity range (from 1 to 100 mm/h). structure.
Even for specific areas, such as Montreal’s downtown When applying this framework to outputs from
core (Fig. 4c) the performance of the DL framework is coarser climate simulations in future studies, it is
quite adequate and much better than the original LR important to account for the fact that the actual reso-
data. lution of the precipitation data is coarser than its grid
spacing, meaning that the data will be overly smooth
Fig. 4 a Ratio of spatial frequency of rainfall intensity above 50 mm/h with respect to HR truth, from LR GEM (left) and DL model (right). b
Frequency–intensity relationships for the three datasets in a, over the entire HR domain. c As b, but for Montreal’s downtown core
Teufel et al. Geoscience Letters (2023) 10:19 Page 8 of 9
and preprocessing (e.g., sharpening) would be required Received: 29 July 2022 Accepted: 28 March 2023
to increase the quality and realism of the estimates
produced by the framework. Potential future improve-
ments to this framework include taking into account
the spatial structure of precipitation and/or control- References
Barrera-Animas AY, Oyedele LO, Bilal M, Akinosho TD, Delgado JMD, Akanbi
ling variables (e.g., buoyancy, wind speed/shear), and LA (2022) Rainfall prediction: a comparative analysis of modern machine
improvements to storm motion estimates, given that learning algorithms for time-series forecasting. Mach Learn Appl
storms often influence atmospheric circulation at local 7:100204. [Link]
Brenowitz ND, Bretherton CS (2018) Prognostic validation of a neural network
scales and thus not always follow regional-scale winds. unified physics parameterization. Geophys Res Lett 45(12):6289–6298.
We intend to explore such ideas in future research. [Link]
Finally, the benefits of a properly trained deep learn- Caballero J, Ledig C, Aitken A, Acosta A, Totz J, Wang Z, Shi W (21–26 July 2017)
Real-time video super-resolution with spatio-temporal networks and
ing framework capable of generating engineering motion compensation. Paper presented at the 2017 IEEE conference on
scale precipitation information cannot be overstated. computer vision and pattern recognition (CVPR)
For a fraction of the computational cost of conven- Chung H, Shin K-S (2018) Genetic algorithm-optimized long short-term
memory network for stock market prediction. Sustainability 10(10):3765
tional methods that involve high-resolution numerical Côté J, Gravel S, Méthot A, Patoine A, Roch M, Staniforth A (1998) The
regional climate models, long simulations of precipita- operational CMC–MRB global environmental multiscale (GEM) model.
tion with ample spatial and temporal detail can readily Part I: design considerations and formulation. Monthly Weather Rev
126(6):1373–1395. [Link]
be performed, and even generating several ensembles tocmge>2.0.co;2
of simulations is readily feasible, which is fundamental Daniel M, Lemonsu A, Déqué M, Somot S, Alias A, Masson V (2019) Benefits of
for quantifying uncertainty in climate projections. This explicit urban parameterization in regional climate modeling to study
climate and city interactions. Clim Dyn 52(5):2745–2764. [Link]
advancement in modeling intense precipitation events 10.1007/s00382-018-4289-x
using physics-informed deep learning framework will Ding Y, Zhu Y, Feng J, Zhang P, Cheng Z (2020) Interpretable spatio-temporal
facilitate the development of effective adaptation and attention LSTM model for flood forecasting. Neurocomputing 403:348–
359. [Link]
mitigation strategies for the climate challenges being Diro GT, Sushama L (2019) Simulating Canadian arctic climate at convection-
faced by highly interconnected engineering systems permitting resolution. Atmosphere 10(8):430
in complex urban environments. It is hoped that the Dong C, Loy CC, He K, Tang X (2016) Image super-resolution using deep con-
volutional networks. IEEE Trans Pattern Anal Mach Intell 38(2):295–307.
outcomes of this study and the developed theoreti- [Link]
cal framework will trigger several additional studies in Girard C, Plante A, Desgagné M, McTaggart-Cowan R, Côté J, Charron M, Gravel
other urban regions of the world. S, Lee V, Patoine A, Qaddouri A, Roch M, Spacek L, Tanguay M, Vaillancourt
PA, Zadra A (2014) Staggered vertical discretization of the Canadian
Acknowledgements environmental multiscale (GEM) model using a coordinate of the log-
The GEM simulations and DL experiments considered in this study were hydrostatic-pressure type. Mon Weather Rev 142(3):1183–1196. [Link]
performed on the supercomputer managed by the Digital Research Alliance doi.org/10.1175/mwr-d-13-00255.1
of Canada and Calcul Québec. The authors would like to acknowledge the two Girihagama L, Naveed Khaliq M, Lamontagne P, Perdikaris J, Roy R, Sushama
anonymous reviewers for their helpful comments. L, Elshorbagy A (2022) Streamflow modelling and forecasting for
Canadian watersheds using LSTM networks with attention mechanism.
Author contributions Neural Comput Appl 34(22):19995–20015. [Link]
BT, FC and LSus designed the deep learning framework in consultation with s00521-022-07523-8
LSun and MNK. BT performed the climate simulations and FC coded the deep Hersbach H, Bell B, Berrisford P, Hirahara S, Horányi A, Muñoz-Sabater J, Nicolas
learning framework. Both FC and BT performed deep learning experiments J, Peubey C, Radu R, Schepers D, Simmons A, Soci C, Abdalla S, Abellan
and analyzed the results. BT drafted the initial manuscript, to which all authors X, Balsamo G, Bechtold P, Biavati G, Bidlot J, Bonavita M, De Chiara G,
contributed substantial revisions, following which they read and approved the Dahlgren P, Dee D, Diamantakis M, Dragani R, Flemming J, Forbes R,
final manuscript. Fuentes M, Geer A, Haimberger L, Healy S, Hogan RJ, Hólm E, Janisková
M, Keeley S, Laloyaux P, Lopez P, Lupu C, Radnoti G, de Rosnay P, Rozum I,
Funding Vamborg F, Villaume S, Thépaut J-N (2020) The ERA5 global reanalysis. Q J
This research was funded by the National Research Council of Canada (New R Meteorol Soc 146(730):1999–2049. [Link]
Beginnings Initiative), Canadian Space Agency (Grant 21SUESDFIM), Natural Huszar P, Halenka T, Belda M, Zak M, Sindelarova K, Miksovsky J (2014) Regional
Sciences and Engineering Research Council of Canada, Trottier Institute for climate model assessment of the urban land-surface forcing over central
Sustainability in Engineering and Design and McGill Sustainability Systems Europe. Atmos Chem Phys 14(22):12393–12413. [Link]
Initiative. acp-14-12393-2014
IPCC (2013) Climate change 2013: the physical science basis. In: Stocker TF,
Availability of data and materials Qin D, Plattner G-K, Tignor M, Allen SK, Boschung J, Nauels A, Xia Y, Bex V,
All data used in this study can be accessed at: [Link] Midgley PM (eds) Contribution of working group I to the 5th assessment
6631996 report of the intergovernmental panel on climate change. Cambridge
University Press, Cambridge
Kendon EJ, Ban N, Roberts NM, Fowler HJ, Roberts MJ, Chan SC, Evans JP,
Declarations Fosser G, Wilkinson JM (2017) Do convection-permitting regional climate
models improve projections of future precipitation change? Bull Am
Competing interests Meteor Soc 98(1):79–93. [Link]
The authors declare that they have no competing interests. Lai WS, Huang JB, Ahuja N, Yang MH (21–26 July 2017) Deep Laplacian pyra-
mid networks for fast and accurate super-resolution. Paper presented at
Teufel et al. Geoscience Letters (2023) 10:19 Page 9 of 9
the 2017 IEEE conference on computer vision and pattern recognition and data mining, Halifax, NS, Canada. [Link]
(CVPR) 3098004
Li L, Yu Q, Yuan Y, Shang Y, Lu H, Sun X (2009) Super-resolution reconstruction Verseghy DL (2011) CLASS—the Canadian land surface scheme (version 3.5),
and higher-degree function deformation model based matching for technical documentation (version 1). Climate Research Division, Science
Chang’E-1 lunar images. Sci China Ser Technol Sci 52(12):3468. [Link] and Technology Branch, Environment Canada
org/10.1007/s11431-009-0334-7 Wang Z, Liu D, Yang J, Han W, Huang T (7–13 Dec 2015) Deep networks for
Liu D, Wang Z, Fan Y, Liu X, Wang Z, Chang S, Huang T (22–29 Oct 2017) Robust image super-resolution with sparse prior. Paper presented at the 2015
video super-resolution with learned temporal dynamics. Paper presented IEEE international conference on computer vision (ICCV)
at the 2017 IEEE international conference on computer vision (ICCV) Wu Y, Teufel B, Sushama L, Belair S, Sun L (2021) Deep learning-based super-
Makansi O, Ilg E, Brox T (2017) End-to-end learning of video super-resolution resolution climate simulator–emulator framework for urban heat studies.
with motion compensation, Cham Geophys Res Lett 48(19):e2021GL094737. [Link]
Masson V (2000) A physically-based scheme for the urban energy budget in L094737
atmospheric models. Bound-Layer Meteorol 94(3):357–397. [Link] Xie Y, Franz E, Chu M, Thuerey N (2018) TempoGAN: a temporally coher-
org/10.1023/A:1002463829265 ent, volumetric GAN for super-resolution fluid flow. ACM Trans Graph
Milbrandt JA, Yau MK (2005) A multimoment bulk microphysics parameteriza- 37(4):Article 95. [Link]
tion. Part I: analysis of the role of the spectral shape parameter. J Atmos Zhang Y, Tian Y, Kong Y, Zhong B, Fu Y (18–23 June 2018) Residual dense net-
Sci 62(9):3051–3064. [Link] work for image super-resolution. Paper presented at the 2018 IEEE/CVF
Mölders N, Olson MA (2004) Impact of urban effects on precipitation in high conference on computer vision and pattern recognition
latitudes. J Hydrometeorol 5(3):409–429. [Link]
7541(2004)005%3c0409:ioueop%3e2.0.co;2
Oke TR (1982) The energetic basis of the urban heat island. Q J R Meteorol Soc Publisher’s Note
108(455):1–24. [Link] Springer Nature remains neutral with regard to jurisdictional claims in pub-
Pradhan P, Tingsanchali T, Shrestha S (2020) Evaluation of soil and water assess- lished maps and institutional affiliations.
ment tool and artificial neural network models for hydrologic simulation
in different climatic regions of Asia. Sci Total Environ 701:134308. [Link]
doi.org/10.1016/j.scitotenv.2019.134308
Prein AF, Langhans W, Fosser G, Ferrone A, Ban N, Goergen K, Keller M, Tölle M,
Gutjahr O, Feser F, Brisson E, Kollet S, Schmidli J, van Lipzig NPM, Leung
R (2015) A review on regional convection-permitting climate modeling:
demonstrations, prospects, and challenges. Rev Geophys 53(2):323–361.
[Link]
Ray SN, Chattopadhyay S (2021) Analyzing surface air temperature and rainfall
in univariate framework, quantifying uncertainty through Shannon
entropy and prediction through artificial neural network. Earth Sci Inf
14(1):485–503. [Link]
Reichstein M, Camps-Valls G, Stevens B, Jung M, Denzler J, Carvalhais N,
Prabhat (2019) Deep learning and process understanding for data-driven
Earth system science. Nature 566(7743):195–204. [Link]
s41586-019-0912-1
Sajjadi MSM, Vemulapalli R, Brown M (18–23 June 2018) Frame-recurrent video
super-resolution. Paper presented at the 2018 IEEE/CVF conference on
computer vision and pattern recognition
Shepherd JM, Burian SJ (2003) Detection of urban-induced rainfall anomalies
in a major coastal city. Earth Interact 7(4):1–17. [Link]
1087-3562(2003)007%3c0001:douira%3e2.0.co;2
Stengel K, Glaws A, Hettinger D, King RN (2020) Adversarial super-resolution
of climatological wind and solar data. Proc Natl Acad Sci 117(29):16805–
16815. [Link]
Tao X, Gao H, Liao R, Wang J, Jia J (22–29 Oct 2017) Detail-revealing deep
video super-resolution. Paper presented at the 2017 IEEE international
conference on computer vision (ICCV)
Teufel B, Sushama L (2022) High-resolution modelling of climatic hazards
relevant for Canada’s northern transportation sector. Clim Dyn. [Link]
doi.org/10.1007/s00382-022-06265-6
Teufel B, Sushama L, Poitras V, Dukhan T, Bélair S, Miranda-Moreno L, Sun L,
Sasmito AP, Bitsuamlak G (2021) Impact of COVID-19-related traffic slow-
down on urban heat characteristics. Atmosphere 12(2):243
Trinh D, Luong M, Dibos F, Rocchisani J, Pham C, Nguyen TQ (2014) Novel
example-based method for super-resolution and denoising of medical
images. IEEE Trans Image Process 23(4):1882–1895. [Link]
1109/TIP.2014.2308422
UN (2018) World urbanization prospects: the 2018 revision. Department of
Economic and Social Affairs, United Nations New York, NY, USA
Van SP, Le HM, Thanh DV, Dang TD, Loc HH, Anh DT (2020) Deep learning
convolutional neural network in rainfall–runoff modelling. J Hydroinf
22(3):541–561. [Link]
Vandal T, Kodra E, Ganguly S, Michaelis A, Nemani R, Ganguly AR (2017)
DeepSD: generating high resolution climate change projections through
single image super-resolution. Paper presented at the proceedings of
the 23rd ACM SIGKDD international conference on knowledge discovery