Radar-Based Volumetric Precipitation Nowcasting: A 3D Convolutional Neural Network with U-Net Architecture Peter Pavlík1,2 , Viera Rozinajová2,3 and Anna Bou Ezzeddine2 1 Faculty of Information Technology, Brno University of Technology, Božetěchova 1/2, Brno-Královo Pole, 612 00, Czechia 2 Kempelen Institute of Intelligent Technologies, Mlynské Nivy II. 18890/5, Bratislava, 821 09, Slovakia 3 Slovak Centre for Research of Artificial Intelligence - slovak.AI, Slovakia Abstract In recent years – like in many other domains – deep learning models have found their place in the domain of precipitation nowcasting. Many of these models are based on the U-Net architecture, which was originally developed for biomedical segmentation, but is also useful for the generation of short-term forecasts and therefore applicable in the weather nowcasting domain. The existing U-Net-based models use sequential radar data mapped into a 2-dimensional Cartesian grid as input and output. We propose to incorporate a third - vertical - dimension to better predict precipitation phenomena such as convective rainfall and present our results here. We compare the nowcasting performance of two comparable U-Net models trained on two-dimensional and three-dimensional radar observation data. We show that using volumetric data results in a small, but significant reduction in prediction error. Keywords precipitation nowcasting, radar imaging, U-Net 1. Introduction systems because it requires highly accurate and con- stantly updated data about precipitation fields, i.e. the Accurate precipitation nowcasting is important for plan- location of storms, wind, fog, snow etc. Weather radar ning various human activities and tasks such as agri- systems are essential for nowcasting because they di- culture, construction building or winter road mainte- rectly observe precipitation particles with an update rate nance. Nowcasting is defined by the World Meteoro- of a few minutes [1]. See Figure 1 for an example of a logical Agency as forecasting with local detail, by any radar precipitation map. method, over a period from the present to six hours In the last few years, deep learning precipitation now- ahead, including a detailed description of the present casting approaches, such as convolutional neural net- weather [1]. works (CNN), started to gain attention. From the initial In practice, simpler - and therefore faster - models out- ConvLSTM model [2], through encoder-decoder U-Net perform complex Numerical Weather Prediction (NWP) architectures [3, 4], to the recently-introduced GAN- models at the task of precipitation nowcasting because based approaches [5, 6], the CNN models proved to NWP models cannot consider the latest observations due consistently outperform the operational state-of-the-art to their long inference time. The highly sophisticated methods in the domain [6]. NWP models usually need hours to produce their fore- Most precipitation nowcasting models only use the casts and so they are not able to take into consideration radar data mapped to a 2D Cartesian grid, aggregating the latest data observations. Even a simple model that the vertical dimension, even though the raw output of can quickly output a prediction will outperform the NWP weather radar systems consists of multiple measurements models at the task of precipitation nowcasting simply by at different elevation angles and polar coordinates that the fact that it can consider the present data. Nowcast- capture the precipitation phenomena in 3-dimensional ing models can work in conjunction with NWP models space around the radar. and use their long-term forecasts as additional inputs to We propose using volumetric data from multiple alti- further refine their nowcasts [1]. tudes to give the model as much data about the observa- Precipitation nowcasting is usually performed using tion as possible. Providing information about the vertical temporal extrapolation of past data from weather radar motion of precipitation particles, as well as their vertical CDCEO 2022: 2nd Workshop on Complex Data Challenges in Earth extension, could potentially be valuable for the model, as Observation, July 25, 2022, Vienna, Austria they are an important factor in predicting the behavior Envelope-Open peter.pavlik@kinit.sk (P. Pavlík); viera.rozinajova@kinit.sk of convective storms [7]. (V. Rozinajová); anna.bou.ezzeddine@kinit.sk (A. B. Ezzeddine) We compare two models - a reference U-Net architec- Orcid 0000-0002-7468-5503 (P. Pavlík); 0000-0003-1302-6261 ture based on existing research [3, 4] and an alternative (V. Rozinajová); 0000-0002-3341-6059 (A. B. Ezzeddine) © 2021 Copyright for this paper by its authors. Use permitted under Creative with 2D convolutional layers replaced by 3D convolution. Commons License Attribution 4.0 International (CC BY 4.0). CEUR Workshop Proceedings http://ceur-ws.org ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org) We evaluate their performance in the task of predicting 50 The first deep learning approach applied to the task of 300 40 precipitation nowcasting was a ConvLSTM model pre- sented in [2] that outperformed the operational optical- 250 flow-based ROVER nowcasting system. Experiments 30 with other CNN architectures started, such as a Con- 200 vGRU model from [15] or a U-Net-based architecture 20 introduced in [16]. The U-Net architectures, originally de- dBZ 150 veloped for segmentation of medical images [17], proved 10 to be quite popular with models such as RainNet[3] and 100 SmaAt-U-Net[4] further exploring this approach. 0 The previously mentioned neural network regression 50 models trying to nowcast the future state of precipita- 10 tion fields were affected by blurring. When using tra- 0 0 50 100 150 200 250 300 ditional gridpoint-based verification statistics such as Mean Squared Error (MSE) as the training loss function, Figure 1: A single radar echo observation. The shown re- we face the so-called “double penalty problem”. A fore- flectivity values represent reflectivity captured at 2 km above cast of a precipitation feature that is correct in terms of radar (CAPPI). The reflectivity map is overlaid over a satel- lite image of the appropriate area centered on the Malý Ja- intensity, size, and timing, but incorrect concerning loca- vorník radar station generated using Google Earth Engine [8]. tion, results in very large mean square error [18]. This Landsat-8 image courtesy of the U.S. Geological Survey. causes the model to produce blurry outputs to mitigate the penalisation caused by spatially incorrect precipita- tion features. The blurry predictions pose one of the biggest chal- a single constant-altitude radar reflectivity observation lenges for anyone trying to develop a nowcasting model 30 minutes into the future. based on machine learning as such predictions have diffi- Our experiments show that providing volumetric data culties predicting extreme events due to the smoothing. from multiple altitude levels results in small, but statisti- Recently, this problem started to be addressed by training cally significant reduction of prediction error. models using the Generative Adversarial Network (GAN) approach, the most prominent being DGMR[6]. They 2. Related Work introduced a GAN framework[19] to solve the problem of blurry predictions present in other deep learning pre- Many automated nowcasting systems that employ var- cipitation nowcasting models such as RainNet. Model ious inputs and computation approaches are in use to- is trained using a combination of two discriminators in- day [9, 10, 11, 12, 13]. These systems are generally based spired by existing research in video generation and a on extrapolating past observed rainfall data forwards in regularization term that comprise the loss function. The time. They typically estimate the future advection based first discriminator, spatial, discourages blurry predictions on motion observed in the most recent radar images us- while the second one, temporal, discourages jumpy pre- ing cross-correlation or optical flow techniques [1]. dictions. The regularization term penalizes deviations Some nowcasting systems use the cell tracking ap- between the observed radar sequences and the model proach. They firstly identify storms in the radar scan prediction. The DGMR model can be currently consid- and then locate the corresponding object in the consecu- ered the state-of-the-art in the precipitation nowcasting tive scans to track its motion. Cell tracking is useful for domain. tracking severe storms and is useful for generating early warnings [1]. 2.1. Motivation for Volumetric The shortcoming of these advection nowcasting meth- Nowcasting ods is the assumption that the observed precipitation field will not change, only move elsewhere. Therefore, The application of deep learning models for precipitation they lack the capability to predict beginning of new pre- nowcasting is the focus of many research works. How- cipitation phenomena such as convective initiation (start ever, the vast majority of the models use 2-dimensional of a storm triggered by rising moist warm air) or the aggregate radar products and thus throw away any infor- decaying of the storm at the end of its lifecycle [1, 14]. mation which can be gained from processing the vertical In the past years, data-driven approaches using deep structure of precipitation objects captured by the radar. learning to construct precipitation nowcasting models When reviewing the existing works in the precipi- to mitigate these limitations have started to gain atten- tation nowcasting domain, we identified a need to ex- tion [2, 3, 6]. plore the effect of working with 3-dimensional volumetric 45.0 Deg. 2017-08-06T13:00:06Z radar data. By processing the data into a 2D aggregated 10 Equivalent reflectivity factor 60 map, we lose all information about the vertical structure of the precipitation particles detected by the radar. The 8 equivalent reflectivity factor (dBZ) 40 Distance Above radar (km) model trained in this way cannot consider the vertical 6 movement of particles caused by updraft or downdraft 20 and predict the future precipitation accordingly. 4 0 Compared to 2-dimensional precipitation nowcasting, volumetric models are much less prevalent. One such 2 20 model was presented in [20], where a ConvLSTM model 0 was used to predict future radar reflectivity. The model 0 50 100 150 Distance from radar (km) 200 input shape is 18×18×20 (18×18 km with 1 km resolution, Figure 2: Vertical slice of a single radar reflectivity observa- 10 km above at 500 m resolution) provided at multiple tion at a set azimuth. The separate ”rays” at different elevation time steps, each one is processed by a 3D-CNN first, angles are identifiable. then passed on to ConvLSTM sequential network. The output is a classification for the central region of 6 × 6 km predicting whether the reflectivity in the next 30 and 60 single radar observation. minutes will exceed a set threshold. The final result is Since the convolutional neural network models cannot a binary map with resolution of 6 × 6 km. The problem process the data in polar coordinates, we need to convert with this approach is that the model cannot consider any them into Cartesian maps. We processed the data using fast moving precipitation particles, since it cannot see the Py-ART Python library [21]. The radar echo obser- more than 6 km past its target region. Also, the target vations are typically aggregated into precipitation maps region size of 6 × 6 km can hardly be considered a high in two forms. The first one is Constant Altitude Plan spatial resolution, which is one of the defining traits of Position Indicator (CAPPI), which displays reflectivity nowcasting. gate values at certain altitude slice above radar. The other One other work worth mentioning is a 3D-CNN+GAN is CMAX, which aggregates the vertical dimension and hybrid model from [5]. This model is quite sophisticated. displays the maximum value in the vertical column for It uses the GAN-based approach to predict plausible data each data point. If a 3D volume is created from multiple and a weighted MSE loss function to give more impor- CAPPI maps at different altitude levels, the product is tance to high reflectivity values, resulting in better ability called MCAPPI. to predict extreme precipitation events and reduce out- The reflectivity maps can be converted to rainfall rate put blurring. However, the third data dimension is not maps using the Marshall-Palmer Z-R relationship[22]: actually the altitude above radar we want to consider, but time - i.e. the past observations are not as separate 𝑍 = 200𝑅1.6 (1) channels, but form a 3D volume. Nevertheless, the model drives the development of 3D-CNN models for precipita- where 𝑍 is the reflectivity factor and 𝑅 is the rainfall tion nowcasting. rate in 𝑚𝑚/ℎ. 3. Radar Reflectivity Dataset 3.1. Training data selection The dataset requires filtering before training since the ma- To explore the effect of volumetric precipitation now- jority of the observations are of clear skies with nothing casting, we collaborated with the Slovak Meteorological to learn from. Most of the observations from the dataset Institute that provided us a dataset of roughly 3.5 years therefore have no value for training the model and could of reflectivity data from Malý Javorník weather radar even negatively affect the training by biasing the outputs station. The data is captured in 5 minute intervals. The toward clear sky prediction, while we are mostly inter- dataset consists of 355 761 separate observations in the ested in non-trivial cases with high precipitation. We ODIM HDF5 format. filtered the images as follows: The radar captures the precipitation particles in the air by measuring returned radar wave power (echo) after 1. Create a CAPPI radar reflectivity map at 2 km hitting precipitation particles. This value is called reflec- altitude above radar at 1 × 1 km resolution and tivity, measured in logarithmic dimensionless units called select a center slice of size 336 × 336 km. decibels (dBZ). The data consists of reflectivity values 2. Convert reflectivity to rainfall rate according to at the so-called reflectivity gates in multiple elevation Marshall-Palmer Z-R relationship (1). angles distributed around the radar station and encoded 3. Compute the ratio of rainy to clear pixels (thresh- in polar coordinates. See Figure 2 for a vertical slice of a old 0.05 mm/5 min or 0.6 mm/h - corresponds to slight rain). 4. If the rainfall map contains at least 20% of rainy Set No. of obs. % of original pixels and 11 previous observations are available, Full Dataset 355761 100 add it to the target observation set. Target Observations 9018 2.53 Target + Lead Obs. 11310 3.18 Each selected target observation was included in the Training Set Targets 6515 1.83 training dataset, along with a set number of previous ob- Validation Set Targets 1150 0.32 servations to serve as inputs and non-target intermediary Test Set Targets 1353 0.38 outputs. For our models, we decided to use 6 observa- tions as input and 6 as output, effectively predicting the Table 1 precipitation half an hour in advance based on the last The observation count of the full dataset, the subset selected for training according to the training data selection described half hour of data. This means that for each target obser- in Section 3.1 and the sizes of train, test, validation splits. vation, we also needed to include 11 leading observations in the dataset. This process returned 9 018 suitable tar- get images which together with the necessary leading images represent 3.18% of the original dataset. data as much as possible. The GPU processing time (dis- It should be noted that the data converted to rainfall regarding the time to move the data to memory) was not described above was not used for training, only for fil- affected, with both models needing around 6 ms of GPU tering the target observations based on the ratio of rainy time to generate a single output on our hardware. pixels. The actual training data used reflectivity directly for both 2D images and 3D volumes. The 2D dataset 4.1. Training and Evaluation was a collection of CAPPI radar reflectivity maps at 2 km altitude above radar. A 3D dataset was a collection of To train and evaluate the models, the training dataset CAPPI radar reflectivity maps at 8 altitude levels above was split into training, validation and test subsets in radar, from 500 m.a.r to 4000 m.a.r. The extent of the chronological order. The last 15% of target observations data was set to 336 × 336 km centered on the radar sta- were selected for the test set, the rest was chosen for tion with spatial resolution of 1 × 1 km for both 2D and training. Out of these, the last 15% of target observations 3D data, resulting in images of size 336 × 336 pixels and were again selected for validation and the rest was used 8 × 336 × 336 voxels respectively for a single observation. as training samples. See Table 1 for the exact number of observations in each set. Adam optimizer was used for training the model. To 4. Model Architectures find the optimal training model hyperparameters - start- ing learning rate, optimizer learning rate scheduler pa- To compare the impact of adding a vertical dimension as rameters and gradient clipping threshold - we utilized the fairly as possible, we chose a basic U-Net architecture in- Bayesian sweep search provided by Weights & Biases[24]. spired by models developed in [3, 4] as a reference model. We trained 20 models with 2D CNN architecture and 5 As U-Net is a fully convolutional neural network, convert- with 3D CNN architecture. The best performing model of ing it to process volumetric data is a trivial task - mostly each architecture variant was selected for performance just a matter of replacing 2D convolutional layers with evaluation. See Table 2 for all the possible hyperparam- 3D convolutions. Besides this, the model only required eter values and the best performing ones for both 2D replacing 2D max-pooling layers in the encoder for 3D and 3D models. Early stopping after 15 non-improving max-pooling and bilinear upsample in the decoder for epochs was utilized. trilinear. See Figure 3 for the specific number of channels Choosing the right metric to evaluate the performance and kernel sizes at each layer of the model. Both were of precipitation nowcasting models is not simple. The implemented using the PyTorch library [23]. correct method depends on a model’s use-case and no The conversion of the model from 2D to 3D convo- single composite measure is currently able to objectively lutions was mostly straightforward and resulted in in- evaluate performance of precipitation nowcasting mod- creasing the number of trainable parameters 3-fold from els [1]. While we outlined the shortcomings of using roughly 17 to 52 million. The three-fold increase is based MSE to evaluate precipitation nowcasting models above on the fact that the model uses convolution kernels of in Section 2, we are using MSE as the loss function and the size 3 at every convolutional layer, therefore each kernel primary evaluation metric despite the double penaliza- has 27 (3 × 3 × 3) instead of 9 (3 × 3) weights (disregarding tion effect that occurs since it is still the most commonly bias and multiple channels). Other architectural parame- used metric in this domain. Additionally, to provide more ters of the model such as number of kernels at each layer insight into model performance, we are also computing were kept the same for the comparison between these mean model accuracy, precision, recall and F1 scores on models to be fair and dependent solely on the provided binarized precipitation maps using a threshold value of 6 64 64 6 (8x)336x336 128 (8x)168x168 64 256 (8x)84x84 128 512 Double Conv Skip connection (8x)42x42 256 Single Conv Max Pooling (8x)21x21 512 Upsample Figure 3: Diagram of the used U-Net model encoder-decoder architectures and the feed-forward process for both the 2D and 3D variant of the model. Each rectangle represents a multi-channel feature map with the number of channels shown above (or below in the decoder part). The spatial resolution of the feature maps at each level is shown at the left side of the diagram (the vertical dimension size of the 3D model is in brackets). Each arrow represents an operation with the data, see legend at the bottom right. The kernels of the double convolution operation are of size 3 × 3 or 3 × 3 × 3, the kernels of the final single convolution operation are of size 1 × 1 or 1 × 1 × 1 and the kernels of the max pooling operation are of size 2 × 2 or 1 × 2 × 2 for 2D and 3D models respectively. All the convolutional layers used the ReLU activation function. Hyperparameter 2D U-Net 3D U-Net Batch size 32 4 Learning rate 5 × 10−5 , 7.5 × 10−5 , 1 × 10−4 , 2.5 × 10−4 , 5 × 10−4 5 × 10−5 , 7.5 × 10−5 , 1 × 10−4 , 2.5 × 10−4 , 5 × 10−4 Opt. LRS Factor 0.5, 0.7, 0.9 0.5, 0.7, 0.9 Opt. LRS Patience 3, 5, 7 3, 5, 7 Grad. Clip. Thres. 0.2, 1, 5 0.2, 1, 5 Table 2 The hyperparameters values searched through during the training of the models using the Weights & Biases bayesian search. The values used for training the best performing models are in bold. The batch size used was the highest possible based on our GPU memory limit. The optimizer learning rate scheduler parameters are functionally meaningless, as both of the models achieved the best performance before the optimizer was triggered to lower the learning rate. Gradient clipping was added to prevent exploding gradient behavior occurring sometimes when a large starting learning rate was selected. 20 dBZ (corresponding to light rain) to differentiate be- into the future based on past radar reflectivity maps at tween rain and no rain areas. This way, we can evaluate the same altitude. Subsequently, we trained a 3D model only the shape of precipitation features and disregard to predict equivalent 3D reflectivity maps at 8 altitude the intensity, which can serve as another valuable metric.levels based on recent volumetric observation data. To Our experiments have shown that higher threshold val- evaluate which model is better at precipitation nowcast- ues corresponding to extreme precipitation events show ing, we evaluate the prediction error on a single CAPPI larger differences between model metrics during evalua- map at 2 km above radar from the target observation tion, however the informative value would be lower due (nowcast 30 minutes in the future). This can be done to such events occurring only in the small minority of because one slice of the output volume of the 3D model the test set observations. matches the altitude level the 2D model was trained on (2000 m.a.r.). A simple euclidean persistence was used as a bench- 5. 2D vs. 3D: A Comparison mark. This benchmark method simply copies the last input observation as the prediction output. Despite the The impact of providing a vertical dimension to the model method being trivial, the precipitation data is highly de- was evaluated by comparing the error rate when predict- pendent on previous observations and so it provides a ing a single reflectivity map at constant altitude above good performance benchmark. Using this benchmark, radar. We trained the 2D model to output the next CAPPI we can also evaluate the rate of change in the data and radar reflectivity maps at 2 km above radar 30 minutes therefore see how ”difficult” it is to make an accurate Model MSE ↓ MAE ↓ Accuracy ↑ Precision ↑ Recall ↑ F1 ↑ Persistence 55.4110 4.7534 0.8307 0.6529 0.6426 0.6457 2D U-Net 22.6510 3.2623 0.8969 0.8257 0.7282 0.7696 3D U-Net 22.0340 3.2124 0.9000 0.8022 0.7833 0.7894 Table 3 Comparison of model results on the test set for each of the chosen metric scores. The ↓ symbol means it is a lower-is-better score, while the ↑ symbolizes a higher-is-better score. The best result for each score is bolded. Target (t+30 min) Persistence 6. Conclusion Our research shows that providing additional informa- 37.5 tion from multiple altitude levels has the potential to increase the nowcasting accuracy, as compared to the 35.0 currently standard approach of using only 2-dimensional precipitation maps. The improvements in error metrics, 32.5 while not groundbreaking, were statistically significant and show that providing more data is worth it, if we can 30.0 2D-CNN model 3D-CNN model afford the increase in model complexity and training time. Even a small reduction in prediction error can be bene- 27.5 ficial in many applications and our preliminary results show that volumetric nowcasting can have a positive 25.0 impact. Additionally, volumetric nowcasts undoubtedly pro- 22.5 vide more value to the operators of these nowcasting systems. Reflectivity at different altitudes affects the true 20.0 rainfall rate on the ground in different ways, which can- Figure 4: A visual comparison of nowcasts produced by the not be taken into account from simple 2-dimensional models for a random observation from the test set. Upper left precipitation nowcasts. 3-dimensional predictions of fu- image shows the target observation at 2 km CAPPI. Upper ture reflectivity observations can serve as a more valuable right is the benchmark persistence nowcast. Bottom left is input to the consecutive models mapping the observed the reference 2D-CNN U-Net model nowcast. Bottom right is the corresponding slice of our 3D-CNN volumetric U-Net reflectivity to actual the rainfall rate on the ground. model nowcast. While both U-Net models show the expected While the field of precipitation nowcasting using neu- blurring, the volumetric model is affected less, with larger ral networks is not new, there are still more uncertainties areas of high reflectivity (shown in dark red). This is desirable, regarding best practices that should be comprehensively as the model is better at predicting extreme events. explored and compared. There are several open questions to answer, e.g.: Is it better to train the model directly on the captured reflectivity data or the data converted to rainfall rate? How many previous observations should prediction for each sample. be provided to the model? How to convert radar obser- The results in Table 3 show that the best 3D-CNN vations to actual rainfall on the ground as accurately as U-Net model slightly outperformed the best 2D-CNN possible? These are just some of the interesting problems counterpart. On average, the 3D model achieved lower that need to be explored in the future. prediction error on the test set, in both MSE and MAE metrics. The improvement is small, but statistically sig- nificant (paired t-test at 0.99 confidence level on test set Acknowledgments MSE scores rejected the null hypothesis that the means of 2D and 3D model error scores are the same, p-value This research was partially supported by TAILOR, a is very close to zero). The area-based metrics also show project funded by EU Horizon 2020 research and innova- small improvements, with accuracy and F1 scores being tion programme under GA No 952215; by The Ministry slightly higher. Based on considerably higher recall and of Education, Science, Research and Sport of the Slovak lower precision, we can assume the 3D model predicts Republic under the Contract No. 0827/2021; and by Life larger precipitation bodies on average. See Figure 4 for a Defender - Protector of Life, ITMS code: 313010ASQ6, co- visual comparison of the model outputs. financed by the European Regional Development Fund (Operational Programme Integrated Infrastructure). References abilistic precipitation forecasting scheme which merges an extrapolation nowcast with downscaled [1] F. Schmid, Y. Wang, A. Harou, Nowcasting nwp, Quarterly Journal of the Royal Meteorologi- guidelines–a summary, Bulletin nº 68 (2019) 2. cal Society: A journal of the atmospheric sciences, [2] X. Shi, Z. Chen, H. Wang, D.-Y. Yeung, W.-K. Wong, applied meteorology and physical oceanography W.-c. Woo, Convolutional lstm network: A machine 132 (2006) 2127–2155. learning approach for precipitation nowcasting, Ad- [14] S. Pulkkinen, D. Nerini, A. A. Pérez Hortal, vances in neural information processing systems C. Velasco-Forero, A. Seed, U. Germann, L. Foresti, 28 (2015). Pysteps: an open-source python library for proba- [3] G. Ayzel, T. Scheffer, M. Heistermann, Rainnet v1. bilistic precipitation nowcasting (v1.0), Geoscien- 0: a convolutional neural network for radar-based tific Model Development 12 (2019) 4185–4219. URL: precipitation nowcasting, Geoscientific Model De- https://gmd.copernicus.org/articles/12/4185/2019/. velopment 13 (2020) 2631–2644. doi:1 0 . 5 1 9 4 / g m d - 1 2 - 4 1 8 5 - 2 0 1 9 . [4] K. Trebing, T. Stanczyk, S. Mehrkanoon, Smaat- [15] X. Shi, Z. Gao, L. Lausen, H. Wang, D.-Y. Yeung, unet: Precipitation nowcasting using a small W.-k. Wong, W.-c. Woo, Deep learning for precipi- attention-unet architecture, Pattern Recogni- tation nowcasting: A benchmark and a new model, tion Letters 145 (2021) 178–186. URL: https : Advances in neural information processing systems / / www.sciencedirect.com / science / article / pii / 30 (2017). S0167865521000556. doi:h t t p s : / / d o i . o r g / 1 0 . 1 0 1 6 / [16] S. Agrawal, L. Barrington, C. Bromberg, J. Burge, j.patrec.2021.01.036. C. Gazen, J. Hickey, Machine learning for pre- [5] C. Wang, P. Wang, P. Wang, B. Xue, D. Wang, Us- cipitation nowcasting from radar images, CoRR ing conditional generative adversarial 3-d convo- abs/1912.12132 (2019). URL: http://arxiv.org/abs/ lutional neural network for precise radar extrapo- 1912.12132. a r X i v : 1 9 1 2 . 1 2 1 3 2 . lation, IEEE Journal of Selected Topics in Applied [17] O. Ronneberger, P. Fischer, T. Brox, U-net: Convo- Earth Observations and Remote Sensing 14 (2021) lutional networks for biomedical image segmen- 5735–5749. tation, CoRR abs/1505.04597 (2015). URL: http: [6] S. Ravuri, K. Lenc, M. Willson, D. Kangin, R. Lam, //arxiv.org/abs/1505.04597. a r X i v : 1 5 0 5 . 0 4 5 9 7 . P. Mirowski, M. Fitzsimons, M. Athanassiadou, [18] C. Keil, G. C. Craig, A displacement and ampli- S. Kashem, S. Madge, et al., Skilful precipitation tude score employing an optical flow technique, nowcasting using deep generative models of radar, Weather and Forecasting 24 (2009) 1297 – 1308. Nature 597 (2021) 672–677. URL: https://journals.ametsoc.org/view/journals/ [7] C. A. Doswell, Severe convective storms—an wefo / 24 / 5 / 2009waf2222247_1.xml. doi:1 0 . 1 1 7 5 / overview, Severe convective storms (2001) 1–26. 2009WAF2222247.1. [8] USGS/Google, Usgs landsat 8 collec- [19] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, tion 1 tier 1 toa reflectance, 2022. URL: D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio, https : / / developers.google.com / earth - engine / Generative adversarial nets, Advances in neural datasets/catalog/LANDSAT_LC08_C01_T1_TOA. information processing systems 27 (2014). [9] M. Dixon, G. Wiener, Titan: Thunderstorm identifi- [20] W. Zhang, R. Zhang, H. Chen, G. He, Y. Ge, L. Han, cation, tracking, analysis, and nowcasting—a radar- A multi-channel 3d convolutional-recurrent neural based methodology, Journal of atmospheric and network for convective storm nowcasting, in: 2021 oceanic technology 10 (1993) 785–797. IEEE International Geoscience and Remote Sensing [10] A. Hering, C. Morel, G. Galli, S. Sénési, P. Am- Symposium IGARSS, IEEE, 2021, pp. 363–366. brosetti, M. Boscacci, Nowcasting thunderstorms [21] J. J. Helmus, S. M. Collis, The python arm radar in the alpine region using a radar based adaptive toolkit (py-art), a library for working with weather thresholding scheme, in: Proceedings of ERAD, radar data in the python programming language, volume 1, 2004. Journal of Open Research Software 4 (2016). [11] E. Ruzanski, V. Chandrasekar, Y. Wang, The casa [22] J. S. Marshall, W. M. K. Palmer, The distribu- nowcasting system, Journal of Atmospheric and tion of raindrops with size, Journal of Atmo- Oceanic Technology 28 (2011) 640–655. spheric Sciences 5 (1948) 165 – 166. doi:1 0 . 1 1 7 5 / [12] T. Haiden, A. Kann, C. Wittmann, G. Pistotnik, 1520-0469(1948)005<0165:TDORWS>2.0.CO;2. B. Bica, C. Gruber, The integrated nowcast- [23] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Brad- ing through comprehensive analysis (inca) system bury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, and its validation over the eastern alpine region, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, Weather and Forecasting 26 (2011) 166–183. M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, [13] N. E. Bowler, C. E. Pierce, A. W. Seed, Steps: A prob- L. Fang, J. Bai, S. Chintala, Pytorch: An impera- tive style, high-performance deep learning library, in: H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, R. Garnett (Eds.), Advances in Neural Information Processing Systems 32, Cur- ran Associates, Inc., 2019, pp. 8024–8035. URL: http://papers.neurips.cc/paper/9015-pytorch-an- imperative-style-high-performance-deep-learning- library.pdf. [24] L. Biewald, Experiment tracking with weights and biases, 2020. URL: https://www.wandb.com/, soft- ware available from wandb.com.