RainNet: A Large-Scale Dataset for Spatial Precipitation Downscaling
Xuanhong Chen, Kairui Feng, Naiyuan Liu, Yifan Lu, Zhengyan Tong, Bingbing Ni, Ziang Liu, Ning Lin
RRainNet: A Large-Scale Dataset for Spatial Precipitation Downscaling
Xuanhong Chen , * , Kairui Feng , ∗ , Naiyuan Liu , ∗∗ , Yifan Lu , ∗∗ , Zhengyan Tong ,Bingbing Ni , † , Ziang Liu , Ning Lin Shanghai Jiao Tong University, Princeton University, University of Technology Sydney { chen19910528,yifan lu,418004,nibingbing,acemenethil } @sjtu.edu.cn { kairuif,nlin } @princeton.edu , [email protected] Abstract
Spatial Precipitation Downscaling is one of the mostimportant problems in the geo-science community. How-ever, it still remains an unaddressed issue. Deep learn-ing is a promising potential solution for downscaling. Inorder to facilitate the research on precipitation downscal-ing for deep learning, we present the first
REAL (non-simulated) Large-Scale Spatial Precipitation DownscalingDataset,
RainNet , which contains pairs of low-resolution and high-resolution precipitation maps for 17years. Contrary to simulated data, this real dataset cov-ers various types of real meteorological phenomena (e.g.,Hurricane, Squall, etc.), and shows the physical charac-ters -
Temporal Misalignment , Temporal Sparse and
FluidProperties - that challenge the downscaling algorithms. Inorder to fully explore potential downscaling solutions, wepropose an implicit physical estimation framework to learnthe above characteristics. Eight metrics specifically consid-ering the physical property of the data set are raised, whilefourteen models are evaluated on the proposed dataset. Fi-nally, we analyze the effectiveness and feasibility of thesemodels on precipitation downscaling task. The Datasetand Code will be available at https://neuralchen.github.io/RainNet/ .
1. Introduction
The geoscience community’s classical understandingtreats the earth system as a deterministic but extensive phys-ical system, where the law of every particle is pre-known.Following this line, global weather prediction, as a com-putational problem, is comparable to the simulation of thehuman brain and the evolution of the early Universe, and itis performed every day at major operational centers acrossthe world [3]. The improvement of the weather forecast andGeo-data quality saves tremendous money and lives; with * Equal contribution. ∗∗ Equal contribution. † Corresponding author.
Ocean Model ~ ~ ° C0 to 11°C CloudDeepConvection ShallowConvectionWind waves Latentheat flux Sensibleheat flux
Micro-physics (moisture - rain drop)Long-wave flux
Long-waveRadiation Short-waveRadiation
Radiation M i c r o - phy s i c s Surface Process C onv ec ti on Surface Model
Meteorological mechanismPhysical Observation
LR 100km HR 4km O b s e r v a ti on Input Output W ea t h e r F o r eca s t Dynamical Downscaling Statistical Downscaling
Deep Learning
Dyna. Stat. Deep.
Compute Dense Light LightAssumption Yes Yes No NeedConsistency Yes No YesDataset No Need Based on Model
LackDataset!
Figure 1. Downscaling is a vital task in geo-science, which in-volves complex physical processes. Deep learning framework isone potentially promising solution, comparing to computationaldense dynamic methods and spatial-temporal non-consistent sta-tistical methods. No formal dataset challenges the application ofdeep learning in downscaling. In contrast to simulated toy-dataset,our RainNet contains more than low- and correspondinghigh-resolution precipitation map pairs. the fiscal year 2020 budget over $1 billion, NSF funds thou-sands of colleges in U.S. to research on these topics [22].However, in recent years, with computer science develop-ment, a deluge of Earth system data is continuously beingobtained, coming from sensors all over the earth and evenin space. These ever-increasing massive amounts of data,coming in different sources and structures, challenge thegeoscience community, which lacks practical approachesto understand and further utilize the data and a flexibleframework with less pre-knowledge constraints to blend thedata [24] (Shown in Fig. 1).At the same time, deep learning has made an enormousbreakthrough in the field of computer vision, which is ex-1 a r X i v : . [ c s . C V ] D ec remely good at extracting valuable knowledge from nu-merous amounts of data. Witnessing such a situation, sev-eral preliminary works [9, 34, 12] try to introduce machinelearning and deep learning tools to solve meteorologicalproblems, e.g ., spatial precipitation downscaling (increas-ing resolution of initial coarse precipitation dataset, whichplays a vital role in daily weather prediction). However,these methods are only applied to ideal retrospective prob-lems and verified on simulated datasets ( e.g ., bi-cubic ofprecipitation generated by weather forecast model on his-torical events [4]), which significantly weakens the cred-ibility of the feasibility and effectiveness of the methods.To efficiently develop the massive geo-data and speed upthe deep learning research for geosciences, we build thefirst large-scale spatial precipitation downscaling datasetfor deep learning. Furthermore, we propose an implicit dy-namic estimation model as a baseline model to explore po-tential downscaling solutions.The precision of weather prediction is highly dependenton the resolution and reliability of the initial environmen-tal input variables, which drives complex physical weatherprojection models. If the input dataset is nine-times higherresolution, the output will show at least nine-times de-tails. However, the raw data for weather prediction is usu-ally from multi-sources with different levels of confidence.Scientists generally average the raw data to coarse spatialscales ( i.e. , to km ), which are hard to resolve manyfine-scale meteorological phenomena ( e.g ., hurricane, re-gional rainstorm, etc .). The conventional solution for thisissue is Spatial Downscaling , whose main task is to in-fer higher resolution information from raw meteorologicaldata.There are three mainstream spatial downscaling forms:1).
Dynamical Downscaling , it infers the fine-scale pre-cipitation and other climate statues via simulating thefine-scale/regional dynamic process of the coupled land-atmosphere system abstractly described by regional numer-ical weather or physical climate models. Dynamical down-scaling usually demands massive computational resourcesand also needs to set a large number of empirical parame-ters [1], leading to poor timeliness and model performance.2). The second form is
Statistical Downscaling ap-proaches, which aim to establish statistical relationships be-tween small and large-scale meteorological variables. Sincetheir simplicity and low computational cost overhead, thestatistical downscaling forms an attractive alternative to dy-namical one [12]. However, statistical downscaling resultsusually lack of physical logic lies behind different atmo-spheric phenomena in space and time. Simultaneously, thedata contain high uncertainty that impacts the quality ofweather forecasts driven by statistical downscaling meth-ods [16].3). Compared to the above two long-developed forms,
Deep-learning Downscaling is in its infancy, but has shownsuperior performance and reasonable computational re-source overhead. Nevertheless, the existing deep learn-ing methods [9, 34] have NO real and well-organized geo-dataset for training, resulting in the low-confidence modelperformance in real-world scenarios. Furthermore, thesemethods fail to effectively learn spatio-temporal dynamicproperties, which is the most critical and fundamental inspatial downscaling.To overcome the shortages of existing deep-learningdownscaling pipelines, we build the first well-organizedlarge-scale spatial precipitation downscaling dataset fordeep learning, dubbed
RainNet . RainNet provides rainyseason precipitation data for 17 years ( i.e ., ∼ )in the eastern United States, including pairs of low-resolution and high-resolution data. These data are col-lected from satellites, radars and gauge stations, whichcan reveal the multi-source characteristics of meteorologi-cal data. In detail, the multi-source characteristics make themeteorological data show serious Temporal Misalignment and
Temporal Sparse problems, which are more challeng-ing than the image/video super-resolution problem. Un-like video super-resolution, the motion of precipitation re-gion is non-rigid ( i.e ., fluid), while video super-resolutionmainly concerns about rigid body motion estimation. Tofully explore how to alleviate the mentioned predicament,we propose an implicit dynamics estimation downscalingmodel. Our model hierarchically aligns adjacent precipita-tion maps, that is, implicit motion estimation. It is very sim-ple but exhibits very competitive performance. The maincontributions of this paper are:• To the best of our knowledge, we present the firstREAL (non-simulated) Large-Scale Spatial Precipita-tion Donwscaling Dataset for deep learning; Simulta-neously, we deeply analyze the physical characteristicsand challenges of the dataset;• We propose a downscaling model with strong compet-itiveness;• We evaluate 14 competitive potential solutions on theproposed dataset, and analyze the feasibility and effec-tiveness of these solutions.
2. Background
At the beginning of the th century, geoscientists rec-ognized that predicting the state of the atmosphere could betreated as an initial value problem of mathematical physics,wherein future weather is determined by integrating thegoverning partial differential equations, starting from theobserved current weather. Today, this paradigm translatesinto solving a system of nonlinear differential equations atabout half a billion points per time step and accounting2 ep. 13-18, 2018 Hurricane FlorenceSquallHRLRSquall Hurricane
Figure 2.
Dataset Visualization . Please zoom-in the figure for better observation. Please note that the details of the precipitation map arepartially lost due to file compression. Here we plot groups of typical meteorological phenomena (hurricane and squall) in the dataset. Tolearn more about the dataset, please visit our project website https://neuralchen.github.io/RainNet/ . for dynamic, thermodynamic, radiative, and chemical pro-cesses working on scales from hundreds of meters to thou-sands of kilometers and from seconds to weeks [3]. The Navier–Stokes and mass continuity equations (in-cluding the effect of the Earth’s rotation), together with thefirst law of thermodynamics and the ideal gas law, representthe full set of prognostic equations in the atmosphere, de-scribing the change in space and time of wind, pressure,density and temperature is described (shown in Eq. (1–5)) [3].
Momentum Equations: ∂u∂t = − [ u, v, w ] · ∇ u − p ∂p∂x + f v, Zonal ∂v∂t = − [ u, v, w ] · ∇ v − p ∂p∂y − f u, Meridional ∂w∂t = − [ u, v, w ] · ∇ w − p ∂p∂z − g, Vertical (1)
Mass Continuity: ∂ρ∂t = −∇ ([ u, v, w ] · ρ ) , (2) Thermo-dynamic: ∂θ∂t = − [ u, v, w ] · ∇ θ + ˙ Q, (3) Ideal Gas: p = ρRT, (4) Moisture equation: ∂q∂t = − [ u, v, w ] · ∇ q + micro ( q ) . (5)The zonal ( u ), meridional ( v ) and vertical ( w ) windspeed, are driven by air pressure ( p ). The air pressure isdriven by the mass density ( ρ , determined by wind and theingredient of air, e.g ., moisture) and the temperature. Themoisture and air can contain heat, which forms the vari-able latent heat ( θ ). The heat can also come from otherheat sources ( Q ), e.g ., physical processes like sun radia-tion or chemical processes like burning coal. The moisturemay also come from snowmelt or other physical processes,which is described in microphysics ( micro ) - a sub-domainof geoscience.These equations have to be solved numerically usingspatial and temporal discretization because of the mathe-matical intractability of obtaining analytical solutions, andthis approximation creates a distinction between so-calledresolved and unresolved scales of motion. The global weather forecast model, treated as a com-putational problem, relying on high-quality initial data in-3ut. The error of weather forecast would increase expo-nentially over time from this initial error of input dataset.Down-scaling is one of the most important approaches toimprove the initial input quality. Precipitation is one of theessential atmospheric variables that are related to daily life.It could easily be observed, by all means, e.g ., gauge sta-tion, radar, and satellites. Applying down-scaling methodsto precipitation and creating high-resolution rainfall is farmore meaningful than deriving other variables, while it isthe most proper initial task to test deep learning’s power ingeo-science. The traditional down-scaling methods can beseparated into dynamic and statistical down-scaling.Dynamic down-scaling treats the down-scaling as anoptimization problem constraint on the physical law (de-scribed in Eq. (1–5)). The idea of dynamic down-scalingis also a subset of assimilation - the dynamic model meteo-rologists used to blend different data sources, improve dataquality and resolution, and make physical variables con-stant over time and space. In this way, to downscale a 6hours of precipitation data, scientists need to run the physi-cal simulation thousands of times, which would take hourson super-computing centers. This computational difficultyalso limited the data that scientists could utilize to assim-ilate and increase the forecast quality. For down-scaling a7-day precipitation data, people need to simulate the 7-dayprecipitation for thousand times, which would not be ableto finish in days. Thus, this process could not be real-time.As a result, ECMWF(European Centre for Medium-RangeWeather Forecasts, one of the most frontier weather forecastcenters) can only employ the 6-hour earlier data to increaseforecast quality and run four times a day [7].With such a great effort, the dynamic downscaling meth-ods did not reflect the real-world condition. Instead, thesemethods find the most likely precipitation over space andtime under the pre-defined physical law. Though thesephysical laws work fluently on coarse data, cases beyondcurrent knowledge also exist when the observation comesto the more in-depth detail of the atmosphere. This factreminds us that dynamic down-scaling, though works wellunder current weather forecast system, may not help futureforecast system as much as today. A more flexible weatherdown-scaling framework that could easily blend differentsources observations and show the ability to describe morecomplex physical phenomena on different scales is desper-ately in need.Statistical down-scaling is trying to speed up the dy-namic down-scaling process. The input of statistical down-scaling is usually dynamic model results or two differentobservation datasets on different scales. However, due tothe quality of statistical downscaling results, people rarelyapply statistical down-scaling to weather forecasts. Thesemethods are currently applied in the tasks not requiring highdata quality but more qualitative understanding, e.g ., cli- mate projection, which forecasts the weather for hundredsof years on coarse grids and using statistical down-scalingto get detailed knowledge of medium-scale future climatesystem.
3. RainNet: A Large-scale Spatial Precipita-tion Downscaling Dataset
To build up a standard realistic (non-simulated) down-scaling dataset for computer vision, we selected the east-ern coast of the United States, which covers a large region(7 million km ; ◦ ∼ ◦ W , ◦ ∼ ◦ N ) and hasa 20-year high-quality precipitation observations. We col-lected two precipitation data sources from National StageIV QPE Product (StageIV [21]; high resolution at . ◦ (approximately km )) and North American Land Data As-similation System (NLDAS [36]; low resolution at . ◦ (approximately km )). StageIV is mosaicked into a na-tional product at National Centers for Environmental Pre-diction (NCEP), from the regional hourly/6-hourly multi-sensor (radar+gauges) precipitation analyses (MPEs) pro-duced by the 12 River Forecast Centers over the continen-tal United States with some manual quality control doneat the River Forecast Centers (RFCs). NLDAS is con-structed quality-controlled, spatially-and-temporally con-sistent datasets from the gauges and remote sensors to sup-port modeling activities. Both products are hourly updatedand both available from 2002 to the current age.In our dataset, we further selected the Eastern coast re-gion for rain season ( July ∼ N ovember , covering hurri-cane season; hurricanes pour over annual rainfall inless than 10 days). We matched the coordinate system tothe lat-lon system for both products and further labeled allthe hurricane periods happening in the last 17 years. Theseheavy rain events are the largest challenge for weather fore-casting and downscaling products. As a heavy rain couldstimulus a wide-spreading flood, which threatening locallives and arousing public evacuation. If people underes-timate the rainfall, a potential flood would be underrated;while over-estimating the rainfall would lead to unneces-sary evacuation orders and flood protection, which is alsocostly.
At the time of this work, we have collected and pro-cessed precipitation data for the rainy season for 17 yearsfrom 2002 to 2018. One precipitation map pair per hour, precipitation map pairs per day. In detail, we have col-lected months or hours, totaling pairs ofhigh-resolution and low-resolution precipitation maps. Thesize of the high-resolution precipitation map is × ,and the size of the low-resolution is × . Various me-4 RDB C onv U p S a m p li ng C onv RRDB
RRDB …… Shared Weights 𝐈 𝑇 − 𝑁 − 𝐿 Input Frames Implicit dynamic estimation module Downscaling Backbone
Results
L1+Percetual Loss
Ground Truth
Losses A A 𝑁−3 A 𝑁−2 መ 𝐈 𝑇 𝐻 𝐈 𝑇𝐻 …… 𝐈 𝑇 − 𝑁 − + 𝐿 𝐈 𝑇 + 𝑁 − + 𝐿 𝐈 𝑇 + 𝑁 − 𝐿 Figure 3. The pipeline of our proposed baseline model for spatial precipitation downscaling. teorological phenomena and precipitation conditions ( e.g .,hurricanes, squall lines, etc .) are covered in these data.We select typical meteorological phenomena and visu-alize them in Fig. 2. Our data is collected from satellites,radars, gauge stations, etc ., which covers the inherent work-ing characteristics of different meteorological measurementsystems. Compared with traditional methods that generatedata with different resolutions through physical model sim-ulation, our dataset is of great help for deep models to learnreal meteorological laws.
4. Designing Spatial Downscaling Model
Due to the difference between downscaling and tradi-tional figure super-resolution, the metrics work well underSR tasks may not be sufficient for precipitation downscal-ing. Driven by the physical formation of precipitation pro-cess, we propose 6 new metrics: mesoscale peak precipi-tation error (MPPE), heavy rain region error (HRRE), cu-mulative precipitation mean square error (CPMSE), clustermean distance (CMD), heavy rain transition speed (HRTS)and average miss moving degree (AMMD). These 6 met-rics can be separated as reconstruction metrics: MPPE,HRRE, CPMSE, AMMD and dynamic metrics : HRTSand CMD. The MPPE ( mm/hour ) is calculated as the top1/1000 quantile difference between the generated/real rain-fall dataset which considering both spatial and temporalproperty of mesoscale meteorological systems, e.g ., hurri-cane, squall. The HRRE ( km ) measures the difference ofheavy rain coverage on each time slide between generatedand labeled test set, which shows the temporal reconstruc-tion ability of the models. The CPMSE ( mm /hour ) mea-sures the cumulative rainfall difference on each pixel overthe time-axis of the test set, which shows the spatial recon-struction property. The AMMD ( radian ) measures the av-erage angle difference between main rainfall clusters. TheCMD ( km ) compares the location difference of main rain-fall system between the generated and labeled test set.TheHRTS ( km/hour ) measures the difference between main rainfall system moving speed between the generated and la-beled test set which shows the ability for models to capturethe dynamic property. More details about the metrics andtheir equations are given in supplementary materials. Wealso included super-resolution metrics such as LPIPS [40]and PSNR to benchmark the model’s ability in reconstruct-ing realistic rainfall. For more detailed information on met-rics, please see the suppl. In order to help design a more appropriate and effec-tive precipitation downscaling model, we have explored theproperty of the dataset in depth. As mentioned above, ourdataset is collected from multiple sensor sources ( e.g ., satel-lite, weather radar, etc .), which makes the data show a cer-tain extent of misalignment . Our efforts here are not ableto vanquish the misalignment. This is an intrinsic prob-lem brought by the fusion of multi-sensor meteorologicaldata. Limited by observation methods ( e.g ., satellites canonly collect data when they fly over the observation area),meteorological data is usually temporal sparse , e.g ., in ourdataset, the sampling interval between two precipitationmaps is one hour. The temporal sparse leads to serious dif-ficulties in the utilization of precipitation sequences. Ad-ditionally, the movement of the precipitation position is di-rectly related to the cloud. It is a fluid movement processwhich is completely different from the rigid body move-ment concerned in Super-Resolution. At the same time, thecloud will grow or dissipate in the process of flowing, andeven form new clouds, which further complicates the pro-cess. In the nutshell, although existed SR is a potential so-lution for downscaling, there is a big difference betweenthe two. Especially, the three characteristics of downscal-ing mentioned above: Temporal Misalignment , TemporalSparse , Fluid properties , which make the dynamic estima-tion of precipitation more challenging.
As a potential solution,
Super-Resolution (SR) frame-works are generally divided into the Single-Image Super-5 th Channel491 th Channel th Channel
Input
Figure 4. Visualization of the feature map of the precipitation mapprocessed by VGG19. We can see that VGG19 highlights the hur-ricane eye, which illustrates that the perceptual loss is also effec-tive for the precipitation downscaling task.
Resolution (SISR) and the Video Super-Resolution (VSR).Video Super-Resolution is able to leverage multi-frame in-formation to restore images, which better matches the na-ture of downscaling. We will demonstrate this judgmentin Sec. 5.2. The VSR pipeline usually contains threecomponents: deblurring, inter-frame alignment, and super-resolution. Deblurring and inter-frame alignment are im-plemented by the motion estimation module. There are fourmotion estimation frameworks: 1). RNN based [15, 27, 13,11]; 2). Optical Flow [39]; 3). Deformable Convolutionbased [29, 37, 32]; 4). Temporal Concatenation [14, 6, 18].In fact, there is another motion estimation scheme proposedfor the first time in the noise reduction task [28], whichachieves an excellent video noise reduction performance.Inspired by [28], we design an implicit dynamics estima-tion model for the spatial precipitation downscaling. It isworth mentioning that our proposed model and the abovefour frameworks together form a relatively complete candi-date set of dynamic estimation solutions.
Proposed Framework.
As shown in Fig. 3,our framework consists of two components:
Implicitdynamic estimation module and
Downscaling Back-bone . These two parts are trained jointly. Supposethere are N adjacent low-resolution precipitation maps { I LT − N − , .., I LT , ..., I LT + N − } . The task is to reconstructthe high-resolution precipitation map I HT of I LT . The im-plicit dynamic estimation module is composed of multiplevanilla networks A = {A , ..., A N − } sharing weights.Each vanilla network receives three adjacent frames as in-put, outputs and intermediate results. The intermediate re-sult can be considered as a frame with implicit dynamicalignment. We concatenate all the intermediate frames asthe input of the next module. The specific structure of thevanilla network can be found in the supplementary mate-rials. The main task of the downscaling backbone is torestore the high-resolution precipitation map I HT based onthe aligned intermediate frames. In order to make full useof multi-scale information, we use multiple Residual-in-Residual Dense Blocks [33] in the network. We employ theinterpolation+convolution [23] as the up-sampling operatorto reduce the checkerboard artifacts. After processing bydownscaling backbone we get the final estimated HR map ˆI HT . Model objective.
The Downscaling task is essentially torestore high-resolution precipitation maps. We learn fromthe super-resolution task and also apply L1 and perceptualloss as the training loss of our model. The model objectiveis shown below: L ( ˆI HT , I HT ) = (cid:107) ˆI HT − I HT (cid:107) + λ (cid:107) φ ( ˆI HT ) − φ ( I HT ) (cid:107) , (6)where φ denotes the pre-trained VGG19 network [26], weselect the Relu − (without the activator [33]) as the out-put layer. λ is the coefficient to balance the loss terms.
5. Experiments
In order to evaluate the effectiveness of the benchmarkalgorithms, we select the precipitation data from
July to N ovember as the benchmark dataset. The benchmarkcontains pairs of high-resolution and low-resolutionrainfall maps. These data cover various complicated pre-cipitation situations such as hurricanes, squall lines, differ-ent levels of rain, and sunny days. It is sufficient to selectthe rainy season of the year as the test set from the perspec-tive of meteorology, as the climate of one area is normallystable.
The SISR/VSR and the spatial precipitation downscal-ing are similar to some extent, so we argue that the SRmodels can be applied to the task as the benchmark mod-els. The input of SISR is a single image, and the modelinfers a high-resolution image from it. Its main focus is togenerate high-quality texture details to achieve pleasing vi-sual effects. In contrast, VSR models input multiple framesof images ( e.g ., 3 frames, 5 frames, etc .). The core idea ofVSR models is to increase the resolution by complementingtexture information between different frames. It is worthmentioning that VSR models generally are equipped witha motion estimation module to alleviate the challenge ofobject motion to inter-frame information registration. Weevaluated state-of-the-art SISR frameworks ( i.e ., Bicu-bic [15], SRCNN [8], SRGAN [17], EDSR [19], ESR-GAN [33], DBPN [10], RCAN [41]) and VSR frame-works ( i.e ., SRGAN-V, EDSR-V, ESRGAN-V, RBPN [11],EDVR [32]), of which VSR methods ( i.e ., SRGAN-V,EDSR-V, ESRGAN-V) are modified from SISR. In particu-lar, we build SRGAN-V, EDSR-V and ESRGAN-V by con-catenating multiple frames of precipitation maps as the in-put of the model. In addition, we also evaluated the tradi-tional statistics method Kriging, which is widely applied inweather forecasting. The mentioned metrics are used toquantitatively evaluate the performance of these SR models6 R(green) & HR Bicubic SRCNN SRGAN EDSRESRGAN DBPN RCAN SRGAN-V EDSR-VESRGAN-V RBPN EDVR Kriging Ours
Figure 5. Visual comparison with state-of-the-art Super Resolution approaches. Please zoom-in the figure for better observation. Moreresults can be found in suppl.
Approach MPPE ↓ HRRE ↓ AMMD ↓ LPIPS ↓ PSNR ↑ CPMSE ↓ HRTS ↓ CMD ↓ Kriging 0.569 157.716 0.1654 0.0426 49.629 3.590 9.848 11.532Bicubic 1.688 165.408 0.1759 0.045 50.109 2.518 10.585 11.617SRCNN 3.301 166.325 0.1871 0.042 48.266 2.150 10.380 11.590SRGAN 32.34 166.266 0.1834 0.0365 44.361 198.83 8.578 10.439EDSR 2.526 166.243 0.1601 0.0407 50.587 2.180 9.280 11.533ESRGAN 15.218 200.694 0.1915 0.0376 47.876 6.292 8.846 10.568DBPN 3.869 166.340 0.2075 0.0419 48.890 5.3046 9.607 10.994RCAN 2.557 165.803 0.1521 0.0418 50.612 2.102 8.907 11.575SRGAN-V 14.21 160.819 0.1636 0.0342 47.858 50.458 7.531 10.409EDSR-V 2.114 161.048 0.1736 0.0389 50.846 1.9769 7.598 10.470ESRGAN-V 12.623 329.247 0.187 0.0359 48.361 4.498 7.466 9.767RBPN 2.181 160.981 0.1674 0.0351 50.826 1.761 7.768 10.371EDVR 2.113 163.372 0.1684 0.0350 49.826 1.204 8.538 10.627Ours 1.956 158.569 0.1653 0.0311 50.713 1.461 6.821 10.163
Table 1. Comparison with state-of-the-art Super Resolution approaches. The best performance is marked with red (1st best), blue (2ndbest) and green (3rd best). and our method. Further, we select some disastrous weatheras samples for qualitative analysis to test the model’s abilityto learn the dynamic properties of the weather system. Andwe employ the implementation of Pytorch for Bicubic.
We evaluate benchmark frameworks with the benchmarkdataset. The downscaling performances are shown inTab. 1. We divide the indicators mentioned above into twogroups. HRTS and CMD together measure the model’s abil-ity to learn the dynamics of precipitation. MPPE, HRRE,AMMD, LPIPS, PSNR and CPMSE indicators illustrate the7 igure 6. The dynamic property of benchmark algorithms. Theframeworks of VSR are gathered in the lower-left corner of thefigure, which demonstrates that VSR methods are superior to SISRand traditional methods in dynamic properties. model’s ability to reconstruct precipitation. From Tab. 1, wecan learn that the overall performance of the VSR methodsare better than SISR models, which shows that the dynamicproperties mentioned above are extremely important for thedownscaling model. Furthermore, it can be seen from Fig. 6that the SISR method is clustered in the upper right cornerof the scatter plot, and the VSR method is concentrated inthe lower-left corner, which further shows that the dynamicproperties of the VSR methods are overall better than theSISR methods. In addition, our method achieves the st best performance in LIPIPS, HRTS, and achieve the sec-ond best performance on HRRE, CPMSE, CMD. The scoreshows that the implicit dynamic estimation framework usedis feasible and effective. It is worth mentioning that the tra-ditional down-scaling method Kriging has achieved the bestscores in MPPE and HRRE. We visualized the hurricane precipitation map of the th hour in September 2010 and the high-resolution precipi-tation map generated by different methods. As shown inFig. 5, the best perceptual effects are generated by EDVRand Our framework, which is also consistent with the LPIPSscore. Zooming-in the result image, the precipitation mapsgenerated by SRGAN and EDSR present obvious checker-board artifacts. The reason for the checkerboard artifactsshould be the relatively simple and sparse texture patternin precipitation maps. The results generated by Bicubic,RCAN, Kriging and SRCNN are over-smooth. DBPN evencannot reconstruct the eye of the hurricane. Especially, theresult generated by Kriging is as fuzzy as the input LR pre-cipitation map. In conclusion, the visual effects generated by the VSR methods are generally better than the SISRmethods and the traditional method. From the perspectiveof quantitative and qualitative analysis, the dynamics esti-mation framework is very critical for downscaling.
6. Related Works
Downscaling in Geoscience.
Downscaling is a fun-damental task in geoscience and meteorology, which re-searchers have been interested in for a long time [35].Most statistical downscaling methods regard this problemas point-wise regression [20, 30, 25] or direct maximumlikelihood estimation of the high-resolution data from thelow-resolution data [2]. B¨urger et al. [5] compared bias-correction spatial disaggregation (BCSD), quantile regres-sion networks, and expanded downscaling (XSD) for cli-mate downscaling. Four fundamental statistical methodsand three more advanced machine learning methods todownscale daily precipitation in the Northeast United Stateswere compared in [30]. Recently, deep learning methodshave been applied to solve the climate downscaling prob-lem. DeepSD [31] tried to employ SRCNN [8] on precipi-tation downscaling. White et al [34] tested the performanceof Generative Adversarial Network (GAN) on downscal-ing weather models. ClimAlign [9] proposed a novel deeplearning method for statistical downscaling, treating it asan unsupervised domain alignment problem without usingpaired low/high resolution training data.
Single Image Super-Resolution(SISR).
For deep learn-ing based method, Dong et al first proposed an end-to-endmodel SRCNN [8] using deep convolutional network. SR-GAN [17] proposed by Ledig et al uses perceptual loss forfiner texture. The enhanced version ESRGAN [33] intro-duces RRDB as basic network unit, using relative discrimi-nator and perceptual loss of features before activation.
Video Super-Resolution(VSR).
VSR utilizes the tem-poral information of image sequences. Wang et al proposeda framework EDVR [32], which devises PCD module tohandle large motion alignment, and the TSA fusion modulewith temporal and spatial attention. Xie et al proposed thetempoGAN [38] for fluid flow Super-Resolution, synthesiz-ing 4-D physics fields with a temporal discriminator.
7. Conclusion
In this paper, we built the first large-scale real precip-itation downscaling dataset for the deep learning commu-nity. This dataset has 62424 pairs of HR and LR precip-itation maps in total. We believe this dataset will furtheraccelerate the research on precipitation downscaling. Fur-thermore, we analyze the problem in-depth and put forwardthree key challenges: Temporal Misalignment, TemporalSparse, Fluid properties. In addition, we propose an implicitdynamic estimation model to alleviate the above challenges.8t the same time, we evaluated the mainstream SISR andVSR models and found that none of these models can solveRainNet’s problems well. Therefore, the downscaling taskon this dataset is still very challenging.
References [1] T´ercio Ambrizzi, Michelle Reboita, Rosmeri Rocha, andMarta Llopart. The state-of-the-art and fundamental aspectsof regional climate modeling in south america.
Annals of theNew York Academy of Sciences , 1436, 07 2018.[2] Jorge Luis Ba˜no-Medina, Rodrigo Garc´ıa Manzanas,Jos´e Manuel Guti´errez Llorente, et al. Configuration and in-tercomparison of deep learning neural models for statisticaldownscaling. 2020.[3] Peter Bauer, Alan Thorpe, and Gilbert Brunet. Thequiet revolution of numerical weather prediction.
Nature ,525(7567):47–55, 2015.[4] P. Berrisford, D.P. Dee, P. Poli, R. Brugge, Mark Fielding,Manuel Fuentes, P.W. K˚allberg, S. Kobayashi, S. Uppala,and Adrian Simmons. The era-interim archive version 2.0.(1):23, 11 2011.[5] G B¨urger, TQ Murdock, AT Werner, SR Sobie, and AJ Can-non. Downscaling extremes—an intercomparison of multi-ple statistical methods for present climate.
Journal of Cli-mate , 25(12):4366–4388, 2012.[6] Jose Caballero, Christian Ledig, Andrew P. Aitken, Alejan-dro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi.Real-time video super-resolution with spatio-temporal net-works and motion compensation. In , pages 2848–2857.IEEE Computer Society, 2017.[7] PHILIPPE Courtier, J-N Th´epaut, and AnthonyHollingsworth. A strategy for operational implementationof 4d-var, using an incremental approach.
Quarterly Journalof the Royal Meteorological Society , 120(519):1367–1387,1994.[8] Chao Dong, Chen Change Loy, Kaiming He, and XiaoouTang. Image super-resolution using deep convolutional net-works.
IEEE Trans. Pattern Anal. Mach. Intell. , 38(2):295–307, 2016.[9] Brian Groenke, Luke Madaus, and Claire Monteleoni. Cli-malign: Unsupervised statistical downscaling of climatevariables via normalizing flows.
CoRR , abs/2008.04679,2020.[10] Muhammad Haris, Gregory Shakhnarovich, and NorimichiUkita. Deep back-projection networks for super-resolution.In , pages 1664–1673. IEEE Computer Society, 2018.[11] Muhammad Haris, Gregory Shakhnarovich, and NorimichiUkita. Recurrent back-projection network for video super-resolution. In
IEEE Conference on Computer Vision and Pat-tern Recognition, CVPR 2019, Long Beach, CA, USA, June16-20, 2019 , pages 3897–3906. Computer Vision Founda-tion / IEEE, 2019. [12] Xiaogang He, Nathaniel W Chaney, Marc Schleiss, andJustin Sheffield. Spatial downscaling of precipitation us-ing adaptable random forests.
Water resources research ,52(10):8217–8237, 2016.[13] Yan Huang, Wei Wang, and Liang Wang. Bidirectionalrecurrent convolutional networks for multi-frame super-resolution. In
Advances in Neural Information ProcessingSystems , pages 235–243, 2015.[14] Younghyun Jo, Seoung Wug Oh, Jaeyeon Kang, andSeon Joo Kim. Deep video super-resolution network usingdynamic upsampling filters without explicit motion compen-sation. In , pages 3224–3232. IEEE Computer Soci-ety, 2018.[15] Robert Keys. Cubic convolution interpolation for digital im-age processing.
IEEE transactions on acoustics, speech, andsignal processing , 29(6):1153–1160, 1981.[16] Mohammad Sajjad Khan, Paulin Coulibaly, and YonasDibike. Uncertainty analysis of statistical downscaling meth-ods.
Journal of Hydrology , 319(1-4):357–382, 2006.[17] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero,Andrew Cunningham, Alejandro Acosta, Andrew P. Aitken,Alykhan Tejani, Johannes Totz, Zehan Wang, and WenzheShi. Photo-realistic single image super-resolution using agenerative adversarial network. In , pages 105–114. IEEEComputer Society, 2017.[18] Renjie Liao, Xin Tao, Ruiyu Li, Ziyang Ma, and Jiaya Jia.Video super-resolution via deep draft-ensemble learning. In , pages531–539. IEEE Computer Society, 2015.[19] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, andKyoung Mu Lee. Enhanced deep residual networks for sin-gle image super-resolution. In ,pages 1132–1140. IEEE Computer Society, 2017.[20] Mohammad Reza Najafi, Hamid Moradkhani, and Susan AWherry. Statistical downscaling of precipitation using ma-chine learning with optimal predictor selection.
Journal ofHydrologic Engineering , 16(8):650–664, 2011.[21] Brian R. Nelson, Olivier P. Prat, D.-J. Seo, and Emad Habib.Assessment and Implications of NCEP Stage IV Quantita-tive Precipitation Estimates for Product Intercomparisons.
Weather and Forecasting , 31(2):371–394, 02 2016.[22] NSF. Nsf geosciences directorate funding by institution type.
AGI Report , 1(1):1–2, 2020.[23] Augustus Odena, Vincent Dumoulin, and Chris Olah. De-convolution and checkerboard artifacts.
Distill , 1(10):e3,2016.[24] Markus Reichstein, Gustau Camps-Valls, Bjorn Stevens,Martin Jung, Joachim Denzler, Nuno Carvalhais, et al. Deeplearning and process understanding for data-driven earth sys-tem science.
Nature , 566(7743):195–204, 2019.
25] DA Sachindra, Khandakar Ahmed, Md Mamunur Rashid, SShahid, and BJC Perera. Statistical downscaling of precip-itation using machine learning techniques.
Atmospheric re-search , 212:240–258, 2018.[26] Karen Simonyan and Andrew Zisserman. Very deep con-volutional networks for large-scale image recognition. InYoshua Bengio and Yann LeCun, editors, , 2015.[27] Xin Tao, Hongyun Gao, Renjie Liao, Jue Wang, and JiayaJia. Detail-revealing deep video super-resolution. In
IEEEInternational Conference on Computer Vision, ICCV 2017,Venice, Italy, October 22-29, 2017 , pages 4482–4490. IEEEComputer Society, 2017.[28] Matias Tassano, Julie Delon, and Thomas Veit. Fastdvdnet:Towards real-time deep video denoising without flow estima-tion. In , pages 1351–1360. IEEE, 2020.[29] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu.TDAN: temporally-deformable alignment network for videosuper-resolution. In , pages 3357–3366. IEEE, 2020.[30] Thomas Vandal, Evan Kodra, and Auroop R Ganguly. In-tercomparison of machine learning methods for statisticaldownscaling: the case of daily and extreme precipitation.
Theoretical and Applied Climatology , 137(1-2):557–570,2019.[31] Thomas Vandal, Evan Kodra, Sangram Ganguly, AndrewMichaelis, Ramakrishna Nemani, and Auroop R Ganguly.Deepsd: Generating high resolution climate change projec-tions through single image super-resolution. In
Proceedingsof the 23rd acm sigkdd international conference on knowl-edge discovery and data mining , pages 1663–1672, 2017.[32] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, andChen Change Loy. Edvr: Video restoration with enhanceddeformable convolutional networks. In
Proceedings of theIEEE Conference on Computer Vision and Pattern Recogni-tion Workshops , pages 0–0, 2019.[33] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu,Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: En-hanced super-resolution generative adversarial networks. In
Proceedings of the European Conference on Computer Vi-sion (ECCV) , pages 0–0, 2018.[34] BL White, A Singh, and A Albert. Downscaling numericalweather models with gans.
AGUFM , 2019:GC43D–1357,2019.[35] Robert L Wilby, TML Wigley, D Conway, PD Jones, BCHewitson, J Main, and DS Wilks. Statistical downscaling ofgeneral circulation model output: A comparison of methods.
Water resources research , 34(11):2995–3008, 1998.[36] Youlong Xia, Kenneth Mitchell, Michael Ek, JustinSheffield, Brian Cosgrove, Eric Wood, Lifeng Luo, CharlesAlonge, Helin Wei, Jesse Meng, Ben Livneh, Dennis Letten-maier, Victor Koren, Qingyun Duan, Kingtse Mo, Yun Fan, and David Mocko. Continental-scale water and energy fluxanalysis and validation for the north american land data as-similation system project phase 2 (nldas-2): 1. intercompari-son and application of model products.
Journal of Geophys-ical Research: Atmospheres , 117(D3), 2012.[37] Xiaoyu Xiang, Yapeng Tian, Yulun Zhang, Yun Fu, Jan P.Allebach, and Chenliang Xu. Zooming slow-mo: Fast andaccurate one-stage space-time video super-resolution. In , pages 3367–3376. IEEE, 2020.[38] You Xie, Erik Franz, Mengyu Chu, and Nils Thuerey. tem-pogan: A temporally coherent, volumetric gan for super-resolution fluid flow.
ACM Transactions on Graphics (TOG) ,37(4):1–15, 2018.[39] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, andWilliam T. Freeman. Video enhancement with task-orientedflow.
Int. J. Comput. Vis. , 127(8):1106–1125, 2019.[40] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shecht-man, and Oliver Wang. The unreasonable effectiveness ofdeep features as a perceptual metric. In
Proceedings of theIEEE conference on computer vision and pattern recogni-tion , pages 586–595, 2018.[41] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, BinengZhong, and Yun Fu. Image super-resolution using very deepresidual channel attention networks. In Vittorio Ferrari, Mar-tial Hebert, Cristian Sminchisescu, and Yair Weiss, editors,
Computer Vision - ECCV 2018 - 15th European Conference,Munich, Germany, September 8-14, 2018, Proceedings, PartVII , volume 11211 of
Lecture Notes in Computer Science ,pages 294–310. Springer, 2018. ppendix: Supplementary Material In this supplementary material, we illustrate the de-tails of proposed metrics and provide more samples of ourdataset. Furthermore, we discuss the configuration of train-ing and report more experimental results. We will pub-lish the dataset and code on https://neuralchen.github.io/RainNet/ . A. Metrics
Due to the difference between downscaling and tradi-tional figure super-resolution, the metrics work well underSR tasks may not be sufficient for precipitation downscal-ing. Driven by the physical formation of precipitation pro-cess, we propose 6 new metrics: mesoscale peak precipi-tation error (MPPE), heavy rain region error (HRRE), cu-mulative precipitation mean square error (CPMSE), clustermean distance (CMD), heavy rain transition speed (HRTS)and average miss moving degree (AMMD). These 6 met-rics can be separated as reconstruction metrics: MPPE,HRRE, CPMSE, AMMD and dynamic metrics: HRTS andCMD. The MPPE ( mm/hour ) is calculated as the top1/1000 quantile difference between the generated/real rain-fall dataset which considering both spatial and temporalproperty of mesoscale meteorological systems, e.g ., hurri-cane, squall. The HRRE ( km ) measures the difference ofheavy rain coverage on each time slide between generatedand labeled test set, which shows the temporal reconstruc-tion ability of the models. The CPMSE ( mm /hour ) mea-sures the cumulative rainfall difference on each pixel overthe time-axis of the test set, which shows the spatial recon-struction property. The AMMD ( radian ) measures the av-erage angle difference between main rainfall clusters. TheCMD ( km ) compares the location difference of main rain-fall system between the generated and labeled test set. TheHRTS ( km/hour ) measures the difference between mainrainfall system moving speed between the generated and la-beled test set which shows the ability for models to capturethe dynamic property. More details about the metrics andtheir equations are given in supplementary materials.These metrics follows different formulations under thetest set time length T and test set area size A : Mesoscale peak precipitation error (MPPE)
This met-ric measures the ability for the downscaling models to cap-ture the mesoscale peak precipitation. The mesoscale largeweather/meteorological events are happening on a scale of200 km × km , such as hurricane or squall. The ability ofcapturing this metrics would help improve the flood predic-tion, as the precipitation events at this scale could stimuluslarge flooding. By measuring 1/1000 quantile of precipi-tation ( 5000 km ) over temporal and spatial, we couldcapture the precipitation at the mesoscale weather events. Heavy rain region error (HRRE)
This metric measuresthe difference between the reconstructed dataset and thereal high-resolution observations of the heavy rain region.The heavy rain is defined by mm/day , which is a con-ventional benchmark for heavy rain in weather prediction(America: . − . mm/day ; Japan, India and China: − mm/day ; European: − mm/day ). This met-rics is formed by: HRRE = ( 1 T (cid:88) t ( A HR ( P > , t ) − A GT ( P > , t ) ) . , where HR means the high-resolution data and GT is thegenerated data. Cumulative precipitation mean square error (CPMSE)
This metric represents the ability for model to capture thespatial difference of precipitation over a long time, which isusually considered in climatology. Through long time ob-servation, we use this metrics to lay out the impact of missalignment issue and focus on the climatology and spatialrainfall estimation. This metric is formed by:
CP M SE = ( 1 T · A (cid:88) ij ( (cid:88) t P HR ( i, j, t ) − (cid:88) t P GT ( i, j, t )) ) . . Cluster mean distance (CMD)
This metric measures thedistance between the main rainfall clusters between the gen-erated dataset and the high-resolution. This metric blockthe rainfall quantity estimation error and focusing on spa-tial difference on each time slide. For each frame, wefirst calculate out the area size of the heavy rain in HRdataset. We mark the contour of heavy rain in HR dataset as f HR ( x, y, t ) . The area of this contour is marked as A Rain .Then we find the contour of generated rainfall dataset withthe same area size as in f HR ( x, y, t ) and mark the contouras f GT ( x, y, t ) . We calculate out the heavy rain contourdifference between HR and GT dataset under 2-norm. Themetrics could be calculated as: CM D =( (cid:88) t T · A Rain < (cid:90) (cid:90) [ x, y ] f HR ( x, y, t ) − [ x, y ] f GT ( x, y, t ) dxdy > ) . . In which (cid:72) is the area integration; <> means the selfinner product.To further calculate this value, we need to discrete thisvalue as: CM D =( (cid:88) t T · A Rain < (cid:88) i (cid:88) j [ i, j ] f DHR ( i, j, t ) − [ i, j ] f DGT ( i, j, t ) > ) . , f D ∗ becomes 1 when the lattices are on the boundaryof the contours; otherwise it would be 0. Average miss moving degree (AMMD)
These metricsmeasures the ability for model to capture the temporal di-rection of heavy rain. For each frame, we first calculate outthe area size of the heavy rain in HR dataset. We mark thecontour of heavy rain in HR dataset as f HR ( x, y, t ) . Thenfor the last frame in HR dataset, we find the contour of gen-erated rainfall dataset with the same size. We calculate outthe heavy rain contour’s outer normal difference betweenthis and last frame, which has the physical meaning - mov-ing angle of rain clusters. Then we compare the differenceof HR and GT under 2-norm. We also do this for the gener-ated dataset. The metrics could be calculated as: AM M D =( (cid:88) t T · A Rain < (cid:79) t (cid:73) f HR ( x, y, t ) dr HR − (cid:73) f GT ( x, y, t ) · dr GT > ) . . To further calculate this value, we need to discrete thisvalue as:
AM M D =( (cid:88) t T · A Rain < (cid:88) i (cid:88) j ( f OHR ( i, j, t ) − f OHR ( i, j, t − − (cid:88) i (cid:88) j ( f OGT ( i, j, t ) − f OGT ( i, j, t − > ) . , where f O ∗ is the outer normal direction (unitized) of the con-tours. Heavy rain transition speed (HRTS)
These metricsmeasures the ability for model to capture the dynamics(transition speed) of heavy rain. For each frame, wefirst calculate out the area size of the heavy rain in HRdataset. We mark the contour of heavy rain in HR datasetas f HR ( x, y, t ) . Then for the last frame in HR dataset, wefind the contour of generated rainfall dataset with the samesize. We also do this for the generated dataset. We cal-culate out the heavy rain contour difference between thisand last frame. Then we compare the difference of HR andGT under 2-norm. This metrics actually shows the order-1property of dynamics which is shown in main text Eq.1 - thewind blowing effect. The metrics could be calculated as: HRT S =( (cid:88) t T · A Rain < (cid:79) t (cid:90) (cid:90) [ x, y ] f HR ( x, y, t ) − [ x, y ] f GT ( x, y, t ) dxdy > ) . . To further calculate this value, we need to discrete thisvalue as:
HRT S =( (cid:88) t T · A Rain < (cid:88) i (cid:88) j [ i, j ]( f HR ( i, j, t ) − f HR ( i, j, t − − [ i, j ]( f GT ( i, j, t ) − f GT ( i, j, t − > ) . . B. Dataset Details
We show more precipitation maps in proposed dataset.In order to display the dynamic characteristics of the pre-cipitation map more conveniently, we extract the precipita-tion maps of 4 periods and make them into GIFs. TheseGIFs are organized in the attachment of the supplementarymaterials.
C. Extra Results of RainNet
C.1. Detailed Network Structure
The structure of vanilla network in our proposed frame-work is given in Fig. 7. We employ 6 Residual-in-ResidualDense Blocks(RRDB) [33] in our vanilla network.
C.2. Training Details
We select 13 Algorithms as Benchmark: Bicubic [15],SRCNN [8], SRGAN [17], EDSR [19], ESRGAN [33],DBPN [10], RCAN [41], SRGAN-V, EDSR-V, ESRGAN-V, RBPN [11], EDVR [32] and Kriging. These implemen-tations are derived or adapted from publicly available codeprovided by the authors. Since all these methods processthree-channel pictures by default, we modify the number ofinput channels of these models (The precipitation map inour proposed dataset are all single-channel). According toour task, we also adjust the hype parameters of these modelsfor better performance.
C.3. Extra Results
We randomly pick 6 sets of results and show them inFig. 8 ∼
13. In addition, we extract the donwnscaling re-sults (our proposed method) of 5 periods and make theminto GIFs. These GIFs are organized in the attachment ofthe supplementary materials.12
RDB RRDB RRDB C onv C onv C onv C onv Figure 7. The details structure of vanilla network in our proposed model.
HR GT LR GT Bicubic SRCNNDBPNESRGANSRGAN EDSREDVR OursRBPN KrigingRCAN SRGAN-V EDSR-V ESRGAN-V
Figure 8. Visual comparison with state-of-the-art Super Resolution approaches(the specific time: the -th hour in September 2010).Please zoom-in the figure for better observation. Randomly picked results. Please note that the details of the precipitation map are partiallylost due to file compression. R GT LR GT Bicubic SRCNNDBPNESRGANSRGAN EDSREDVR OursRBPN KrigingRCAN SRGAN-V EDSR-V ESRGAN-V
Figure 9. Visual comparison with state-of-the-art Super Resolution approaches(the specific time: the -th hour in November 2011).Please zoom-in the figure for better observation. Randomly picked results. Please note that the details of the precipitation map are partiallylost due to file compression. R GT LR GT Bicubic SRCNNDBPNESRGANSRGAN EDSREDVR OursRBPN KrigingRCAN SRGAN-V EDSR-V ESRGAN-V
Figure 10. Visual comparison with state-of-the-art Super Resolution approaches(the specific time: the -th hour in November 2011).Please zoom-in the figure for better observation. Randomly picked results. Please note that the details of the precipitation map are partiallylost due to file compression. R GT LR GT Bicubic SRCNNDBPNESRGANSRGAN EDSREDVR OursRBPN KrigingRCAN SRGAN-V EDSR-V ESRGAN-V
Figure 11. Visual comparison with state-of-the-art Super Resolution approaches(the specific time: the -th hour in September 2011).Please zoom-in the figure for better observation. Randomly picked results. Please note that the details of the precipitation map are partiallylost due to file compression. R GT LR GT Bicubic SRCNNDBPNESRGANSRGAN EDSREDVR OursRBPN KrigingRCAN SRGAN-V EDSR-V ESRGAN-V
Figure 12. Visual comparison with state-of-the-art Super Resolution approaches(the specific time: the -th hour in August 2011). Pleasezoom-in the figure for better observation. Randomly picked results. Please note that the details of the precipitation map are partially lostdue to file compression. R GT LR GT Bicubic SRCNNDBPNESRGANSRGAN EDSREDVR OursRBPN KrigingRCAN SRGAN-V EDSR-V ESRGAN-V
Figure 13. Visual comparison with state-of-the-art Super Resolution approaches(the specific time: the -th hour in November 2010).Please zoom-in the figure for better observation. Randomly picked results. Please note that the details of the precipitation map are partiallylost due to file compression.-th hour in November 2010).Please zoom-in the figure for better observation. Randomly picked results. Please note that the details of the precipitation map are partiallylost due to file compression.