Search Results
You are looking at 1 - 10 of 24 items for
- Author or Editor: David John Gagne II x
- Refine by Access: All Content x
Abstract
This paper describes the use of convolutional neural nets (CNN), a type of deep learning, to identify fronts in gridded data, followed by a novel postprocessing method that converts probability grids to objects. Synoptic-scale fronts are often associated with extreme weather in the midlatitudes. Predictors are 1000-mb (1 mb = 1 hPa) grids of wind velocity, temperature, specific humidity, wet-bulb potential temperature, and/or geopotential height from the North American Regional Reanalysis. Labels are human-drawn fronts from Weather Prediction Center bulletins. We present two experiments to optimize parameters of the CNN and object conversion. To evaluate our system, we compare the objects (predicted warm and cold fronts) with human-analyzed warm and cold fronts, matching fronts of the same type within a 100- or 250-km neighborhood distance. At 250 km our system obtains a probability of detection of 0.73, success ratio of 0.65 (or false-alarm rate of 0.35), and critical success index of 0.52. These values drastically outperform the baseline, which is a traditional method from numerical frontal analysis. Our system is not intended to replace human meteorologists, but to provide an objective method that can be applied consistently and easily to a large number of cases. Our system could be used, for example, to create climatologies and quantify the spread in forecast frontal properties across members of a numerical weather prediction ensemble.
Abstract
This paper describes the use of convolutional neural nets (CNN), a type of deep learning, to identify fronts in gridded data, followed by a novel postprocessing method that converts probability grids to objects. Synoptic-scale fronts are often associated with extreme weather in the midlatitudes. Predictors are 1000-mb (1 mb = 1 hPa) grids of wind velocity, temperature, specific humidity, wet-bulb potential temperature, and/or geopotential height from the North American Regional Reanalysis. Labels are human-drawn fronts from Weather Prediction Center bulletins. We present two experiments to optimize parameters of the CNN and object conversion. To evaluate our system, we compare the objects (predicted warm and cold fronts) with human-analyzed warm and cold fronts, matching fronts of the same type within a 100- or 250-km neighborhood distance. At 250 km our system obtains a probability of detection of 0.73, success ratio of 0.65 (or false-alarm rate of 0.35), and critical success index of 0.52. These values drastically outperform the baseline, which is a traditional method from numerical frontal analysis. Our system is not intended to replace human meteorologists, but to provide an objective method that can be applied consistently and easily to a large number of cases. Our system could be used, for example, to create climatologies and quantify the spread in forecast frontal properties across members of a numerical weather prediction ensemble.
Abstract
This paper presents an automated approach for classifying storm type from weather radar reflectivity using decision trees. Recent research indicates a strong relationship between storm type (morphology) and severe weather, and such information can aid in the warning process. Furthermore, new adaptive sensing tools, such as the Center for Collaborative Adaptive Sensing of the Atmosphere’s (CASA’s) weather radar, can make use of storm-type information in real time. Given the volume of weather radar data from those tools, manual classification of storms is not possible when dealing with real-time data streams. An automated system can more quickly and efficiently sort through real-time data streams and return value-added output in a form that can be more easily manipulated and understood. The method of storm classification in this paper combines two machine learning techniques: K-means clustering and decision trees. K-means segments the reflectivity data into clusters, and decision trees classify each cluster. The K means was used to separate isolated cells from linear systems. Each cell received labels such as “isolated pulse,” “isolated strong,” or “multicellular.” Linear systems were labeled as “trailing stratiform,” “leading stratiform,” and “parallel stratiform.” The classification scheme was tested using both simulated and observed storms. The simulated training and test datasets came from the Advanced Regional Prediction System (ARPS) simulated reflectivity data, and observed data were collected from composite reflectivity mosaics from the CASA Integrative Project One (IP1) network. The observations from the CASA network showed that the classification scheme is now ready for operational use.
Abstract
This paper presents an automated approach for classifying storm type from weather radar reflectivity using decision trees. Recent research indicates a strong relationship between storm type (morphology) and severe weather, and such information can aid in the warning process. Furthermore, new adaptive sensing tools, such as the Center for Collaborative Adaptive Sensing of the Atmosphere’s (CASA’s) weather radar, can make use of storm-type information in real time. Given the volume of weather radar data from those tools, manual classification of storms is not possible when dealing with real-time data streams. An automated system can more quickly and efficiently sort through real-time data streams and return value-added output in a form that can be more easily manipulated and understood. The method of storm classification in this paper combines two machine learning techniques: K-means clustering and decision trees. K-means segments the reflectivity data into clusters, and decision trees classify each cluster. The K means was used to separate isolated cells from linear systems. Each cell received labels such as “isolated pulse,” “isolated strong,” or “multicellular.” Linear systems were labeled as “trailing stratiform,” “leading stratiform,” and “parallel stratiform.” The classification scheme was tested using both simulated and observed storms. The simulated training and test datasets came from the Advanced Regional Prediction System (ARPS) simulated reflectivity data, and observed data were collected from composite reflectivity mosaics from the CASA Integrative Project One (IP1) network. The observations from the CASA network showed that the classification scheme is now ready for operational use.
Abstract
Probabilistic quantitative precipitation forecasts challenge meteorologists due to the wide variability of precipitation amounts over small areas and their dependence on conditions at multiple spatial and temporal scales. Ensembles of convection-allowing numerical weather prediction models offer a way to produce improved precipitation forecasts and estimates of the forecast uncertainty. These models allow for the prediction of individual convective storms on the model grid, but they often displace the storms in space, time, and intensity, which results in added uncertainty. Machine learning methods can produce calibrated probabilistic forecasts from the raw ensemble data that correct for systemic biases in the ensemble precipitation forecast and incorporate additional uncertainty information from aggregations of the ensemble members and additional model variables. This study utilizes the 2010 Center for Analysis and Prediction of Storms Storm-Scale Ensemble Forecast system and the National Severe Storms Laboratory National Mosaic & Multi-Sensor Quantitative Precipitation Estimate as input data for training logistic regressions and random forests to produce a calibrated probabilistic quantitative precipitation forecast. The reliability and discrimination of the forecasts are compared through verification statistics and a case study.
Abstract
Probabilistic quantitative precipitation forecasts challenge meteorologists due to the wide variability of precipitation amounts over small areas and their dependence on conditions at multiple spatial and temporal scales. Ensembles of convection-allowing numerical weather prediction models offer a way to produce improved precipitation forecasts and estimates of the forecast uncertainty. These models allow for the prediction of individual convective storms on the model grid, but they often displace the storms in space, time, and intensity, which results in added uncertainty. Machine learning methods can produce calibrated probabilistic forecasts from the raw ensemble data that correct for systemic biases in the ensemble precipitation forecast and incorporate additional uncertainty information from aggregations of the ensemble members and additional model variables. This study utilizes the 2010 Center for Analysis and Prediction of Storms Storm-Scale Ensemble Forecast system and the National Severe Storms Laboratory National Mosaic & Multi-Sensor Quantitative Precipitation Estimate as input data for training logistic regressions and random forests to produce a calibrated probabilistic quantitative precipitation forecast. The reliability and discrimination of the forecasts are compared through verification statistics and a case study.
Deep-Learning-Based Precipitation Observation Quality ControlAbstract
We present a novel approach for the automated quality control (QC) of precipitation for a sparse station observation network within the complex terrain of British Columbia, Canada. Our QC approach uses convolutional neural networks (CNNs) to classify bad observation values, incorporating a multiclassifier ensemble to achieve better QC performance. We train CNNs using human QC’d labels from 2016 to 2017 with gridded precipitation and elevation analyses as inputs. Based on the classification evaluation metrics, our QC approach shows reliable and robust performance across different geographical environments (e.g., coastal and inland mountains), with 0.927 area under curve (AUC) and type I/type II error lower than 15%. Based on the saliency-map-based interpretation studies, we explain the success of CNN-based QC by showing that it can capture the precipitation patterns around, and upstream of the station locations. This automated QC approach is an option for eliminating bad observations for various applications, including the preprocessing of training datasets for machine learning. It can be used in conjunction with human QC to improve upon what could be accomplished with either method alone.
Abstract
We present a novel approach for the automated quality control (QC) of precipitation for a sparse station observation network within the complex terrain of British Columbia, Canada. Our QC approach uses convolutional neural networks (CNNs) to classify bad observation values, incorporating a multiclassifier ensemble to achieve better QC performance. We train CNNs using human QC’d labels from 2016 to 2017 with gridded precipitation and elevation analyses as inputs. Based on the classification evaluation metrics, our QC approach shows reliable and robust performance across different geographical environments (e.g., coastal and inland mountains), with 0.927 area under curve (AUC) and type I/type II error lower than 15%. Based on the saliency-map-based interpretation studies, we explain the success of CNN-based QC by showing that it can capture the precipitation patterns around, and upstream of the station locations. This automated QC approach is an option for eliminating bad observations for various applications, including the preprocessing of training datasets for machine learning. It can be used in conjunction with human QC to improve upon what could be accomplished with either method alone.
Abstract
Many statistical downscaling methods require observational inputs and expert knowledge and thus cannot be generalized well across different regions. Convolutional neural networks (CNNs) are deep-learning models that have generalization abilities for various applications. In this research, we modify UNet, a semantic-segmentation CNN, and apply it to the downscaling of daily maximum/minimum 2-m temperature (TMAX/TMIN) over the western continental United States from 0.25° to 4-km grid spacings. We select high-resolution (HR) elevation, low-resolution (LR) elevation, and LR TMAX/TMIN as inputs; train UNet using Parameter–Elevation Regressions on Independent Slopes Model (PRISM) data over the south- and central-western United States from 2015 to 2018; and test it independently over both the training domains and the northwestern United States from 2018 to 2019. We found that the original UNet cannot generate enough fine-grained spatial details when transferred to the new northwestern U.S. domain. In response, we modified the original UNet by assigning an extra HR elevation output branch/loss function and training the modified UNet to reproduce both the supervised HR TMAX/TMIN and the unsupervised HR elevation. This improvement is named “UNet-Autoencoder (AE).” UNet-AE supports semisupervised model fine-tuning for unseen domains and showed better gridpoint-level performance with more than 10% mean absolute error (MAE) reduction relative to the original UNet. On the basis of its performance relative to the 4-km PRISM, UNet-AE is a good option to provide generalizable downscaling for regions that are underrepresented by observations.
Abstract
Many statistical downscaling methods require observational inputs and expert knowledge and thus cannot be generalized well across different regions. Convolutional neural networks (CNNs) are deep-learning models that have generalization abilities for various applications. In this research, we modify UNet, a semantic-segmentation CNN, and apply it to the downscaling of daily maximum/minimum 2-m temperature (TMAX/TMIN) over the western continental United States from 0.25° to 4-km grid spacings. We select high-resolution (HR) elevation, low-resolution (LR) elevation, and LR TMAX/TMIN as inputs; train UNet using Parameter–Elevation Regressions on Independent Slopes Model (PRISM) data over the south- and central-western United States from 2015 to 2018; and test it independently over both the training domains and the northwestern United States from 2018 to 2019. We found that the original UNet cannot generate enough fine-grained spatial details when transferred to the new northwestern U.S. domain. In response, we modified the original UNet by assigning an extra HR elevation output branch/loss function and training the modified UNet to reproduce both the supervised HR TMAX/TMIN and the unsupervised HR elevation. This improvement is named “UNet-Autoencoder (AE).” UNet-AE supports semisupervised model fine-tuning for unseen domains and showed better gridpoint-level performance with more than 10% mean absolute error (MAE) reduction relative to the original UNet. On the basis of its performance relative to the 4-km PRISM, UNet-AE is a good option to provide generalizable downscaling for regions that are underrepresented by observations.
Abstract
Statistical downscaling (SD) derives localized information from larger-scale numerical models. Convolutional neural networks (CNNs) have learning and generalization abilities that can enhance the downscaling of gridded data (Part I of this study experimented with 2-m temperature). In this research, we adapt a semantic-segmentation CNN, called UNet, to the downscaling of daily precipitation in western North America, from the low resolution (LR) of 0.25° to the high resolution (HR) of 4-km grid spacings. We select LR precipitation, HR precipitation climatology, and elevation as inputs; train UNet over the subset of the south- and central-western United States using Parameter–Elevation Regressions on Independent Slopes Model (PRISM) data from 2015 to 2018, and test it independently in all available domains from 2018 to 2019. We proposed an improved version of UNet, which we call Nest-UNet, by adding deep-layer aggregation and nested skip connections. Both the original UNet and Nest-UNet show generalization ability across different regions and outperform the SD baseline (bias-correction spatial disaggregation), with lower downscaling error and more accurate fine-grained textures. Nest-UNet also shares the highest amount of information with station observations and PRISM, indicating good ability to reduce the uncertainty of HR downscaling targets.
Abstract
Statistical downscaling (SD) derives localized information from larger-scale numerical models. Convolutional neural networks (CNNs) have learning and generalization abilities that can enhance the downscaling of gridded data (Part I of this study experimented with 2-m temperature). In this research, we adapt a semantic-segmentation CNN, called UNet, to the downscaling of daily precipitation in western North America, from the low resolution (LR) of 0.25° to the high resolution (HR) of 4-km grid spacings. We select LR precipitation, HR precipitation climatology, and elevation as inputs; train UNet over the subset of the south- and central-western United States using Parameter–Elevation Regressions on Independent Slopes Model (PRISM) data from 2015 to 2018, and test it independently in all available domains from 2018 to 2019. We proposed an improved version of UNet, which we call Nest-UNet, by adding deep-layer aggregation and nested skip connections. Both the original UNet and Nest-UNet show generalization ability across different regions and outperform the SD baseline (bias-correction spatial disaggregation), with lower downscaling error and more accurate fine-grained textures. Nest-UNet also shares the highest amount of information with station observations and PRISM, indicating good ability to reduce the uncertainty of HR downscaling targets.
Abstract
An ensemble postprocessing method is developed for the probabilistic prediction of severe weather (tornadoes, hail, and wind gusts) over the conterminous United States (CONUS). The method combines conditional generative adversarial networks (CGANs), a type of deep generative model, with a convolutional neural network (CNN) to postprocess convection-allowing model (CAM) forecasts. The CGANs are designed to create synthetic ensemble members from deterministic CAM forecasts, and their outputs are processed by the CNN to estimate the probability of severe weather. The method is tested using High-Resolution Rapid Refresh (HRRR) 1–24-h forecasts as inputs and Storm Prediction Center (SPC) severe weather reports as targets. The method produced skillful predictions with up to 20% Brier skill score (BSS) increases compared to other neural-network-based reference methods using a testing dataset of HRRR forecasts in 2021. For the evaluation of uncertainty quantification, the method is overconfident but produces meaningful ensemble spreads that can distinguish good and bad forecasts. The quality of CGAN outputs is also evaluated. Results show that the CGAN outputs behave similarly to a numerical ensemble; they preserved the intervariable correlations and the contribution of influential predictors as in the original HRRR forecasts. This work provides a novel approach to postprocess CAM output using neural networks that can be applied to severe weather prediction.
Significance Statement
We use a new machine learning (ML) technique to generate probabilistic forecasts of convective weather hazards, such as tornadoes and hailstorms, with the output from high-resolution numerical weather model forecasts. The new ML system generates an ensemble of synthetic forecast fields from a single forecast, which are then used to train ML models for convective hazard prediction. Using this ML-generated ensemble for training leads to improvements of 10%–20% in severe weather forecast skills compared to using other ML algorithms that use only output from the single forecast. This work is unique in that it explores the use of ML methods for producing synthetic forecasts of convective storm events and using these to train ML systems for high-impact convective weather prediction.
Abstract
An ensemble postprocessing method is developed for the probabilistic prediction of severe weather (tornadoes, hail, and wind gusts) over the conterminous United States (CONUS). The method combines conditional generative adversarial networks (CGANs), a type of deep generative model, with a convolutional neural network (CNN) to postprocess convection-allowing model (CAM) forecasts. The CGANs are designed to create synthetic ensemble members from deterministic CAM forecasts, and their outputs are processed by the CNN to estimate the probability of severe weather. The method is tested using High-Resolution Rapid Refresh (HRRR) 1–24-h forecasts as inputs and Storm Prediction Center (SPC) severe weather reports as targets. The method produced skillful predictions with up to 20% Brier skill score (BSS) increases compared to other neural-network-based reference methods using a testing dataset of HRRR forecasts in 2021. For the evaluation of uncertainty quantification, the method is overconfident but produces meaningful ensemble spreads that can distinguish good and bad forecasts. The quality of CGAN outputs is also evaluated. Results show that the CGAN outputs behave similarly to a numerical ensemble; they preserved the intervariable correlations and the contribution of influential predictors as in the original HRRR forecasts. This work provides a novel approach to postprocess CAM output using neural networks that can be applied to severe weather prediction.
Significance Statement
We use a new machine learning (ML) technique to generate probabilistic forecasts of convective weather hazards, such as tornadoes and hailstorms, with the output from high-resolution numerical weather model forecasts. The new ML system generates an ensemble of synthetic forecast fields from a single forecast, which are then used to train ML models for convective hazard prediction. Using this ML-generated ensemble for training leads to improvements of 10%–20% in severe weather forecast skills compared to using other ML algorithms that use only output from the single forecast. This work is unique in that it explores the use of ML methods for producing synthetic forecasts of convective storm events and using these to train ML systems for high-impact convective weather prediction.
Abstract
An ensemble precipitation forecast postprocessing method is proposed by hybridizing the analog ensemble (AnEn), minimum divergence Schaake shuffle (MDSS), and convolutional neural network (CNN) methods. This AnEn–CNN hybrid takes the ensemble mean of Global Ensemble Forecast System (GEFS) 3-hourly precipitation forecasts as input and produces bias-corrected, probabilistically calibrated, and physically realistic gridded precipitation forecast sequences out to 7 days. The AnEn–CNN hybrid postprocessing is trained on the European Centre for Medium-Range Weather Forecasts Reanalysis version 5 (ERA5), and verified against station observations across British Columbia (BC), Canada, from 2017 to 2019. The AnEn–CNN hybrid produces more skillful forecasts than a quantile-mapped GEFS baseline and other conventional AnEn methods, with a roughly 10% increase in continuous ranked probability skill score. Further, it outperforms other AnEn methods by 0%–60% in terms of Brier skill score (BSS) for heavy precipitation periods across disparate hydrological regions. Longer forecast lead times exhibit larger performance gains. Verification against 7-day accumulated precipitation totals for heavy precipitation periods also demonstrates that precipitation sequences are realistically reconstructed. Case studies further show that the AnEn–CNN hybrid scheme produces more realistic spatial precipitation patterns and precipitation intensity spectra. This work pioneers the combination of conventional statistical postprocessing and neural networks, and is one of only a few studies pertaining to precipitation ensemble postprocessing in BC.
Abstract
An ensemble precipitation forecast postprocessing method is proposed by hybridizing the analog ensemble (AnEn), minimum divergence Schaake shuffle (MDSS), and convolutional neural network (CNN) methods. This AnEn–CNN hybrid takes the ensemble mean of Global Ensemble Forecast System (GEFS) 3-hourly precipitation forecasts as input and produces bias-corrected, probabilistically calibrated, and physically realistic gridded precipitation forecast sequences out to 7 days. The AnEn–CNN hybrid postprocessing is trained on the European Centre for Medium-Range Weather Forecasts Reanalysis version 5 (ERA5), and verified against station observations across British Columbia (BC), Canada, from 2017 to 2019. The AnEn–CNN hybrid produces more skillful forecasts than a quantile-mapped GEFS baseline and other conventional AnEn methods, with a roughly 10% increase in continuous ranked probability skill score. Further, it outperforms other AnEn methods by 0%–60% in terms of Brier skill score (BSS) for heavy precipitation periods across disparate hydrological regions. Longer forecast lead times exhibit larger performance gains. Verification against 7-day accumulated precipitation totals for heavy precipitation periods also demonstrates that precipitation sequences are realistically reconstructed. Case studies further show that the AnEn–CNN hybrid scheme produces more realistic spatial precipitation patterns and precipitation intensity spectra. This work pioneers the combination of conventional statistical postprocessing and neural networks, and is one of only a few studies pertaining to precipitation ensemble postprocessing in BC.
Abstract
Deep learning models, such as convolutional neural networks, utilize multiple specialized layers to encode spatial patterns at different scales. In this study, deep learning models are compared with standard machine learning approaches on the task of predicting the probability of severe hail based on upper-air dynamic and thermodynamic fields from a convection-allowing numerical weather prediction model. The data for this study come from patches surrounding storms identified in NCAR convection-allowing ensemble runs from 3 May to 3 June 2016. The machine learning models are trained to predict whether the simulated surface hail size from the Thompson hail size diagnostic exceeds 25 mm over the hour following storm detection. A convolutional neural network is compared with logistic regressions using input variables derived from either the spatial means of each field or principal component analysis. The convolutional neural network statistically significantly outperforms all other methods in terms of Brier skill score and area under the receiver operator characteristic curve. Interpretation of the convolutional neural network through feature importance and feature optimization reveals that the network synthesized information about the environment and storm morphology that is consistent with our understanding of hail growth, including large lapse rates and a wind shear profile that favors wide updrafts. Different neurons in the network also record different storm modes, and the magnitude of the output of those neurons is used to analyze the spatiotemporal distributions of different storm modes in the NCAR ensemble.
Abstract
Deep learning models, such as convolutional neural networks, utilize multiple specialized layers to encode spatial patterns at different scales. In this study, deep learning models are compared with standard machine learning approaches on the task of predicting the probability of severe hail based on upper-air dynamic and thermodynamic fields from a convection-allowing numerical weather prediction model. The data for this study come from patches surrounding storms identified in NCAR convection-allowing ensemble runs from 3 May to 3 June 2016. The machine learning models are trained to predict whether the simulated surface hail size from the Thompson hail size diagnostic exceeds 25 mm over the hour following storm detection. A convolutional neural network is compared with logistic regressions using input variables derived from either the spatial means of each field or principal component analysis. The convolutional neural network statistically significantly outperforms all other methods in terms of Brier skill score and area under the receiver operator characteristic curve. Interpretation of the convolutional neural network through feature importance and feature optimization reveals that the network synthesized information about the environment and storm morphology that is consistent with our understanding of hail growth, including large lapse rates and a wind shear profile that favors wide updrafts. Different neurons in the network also record different storm modes, and the magnitude of the output of those neurons is used to analyze the spatiotemporal distributions of different storm modes in the NCAR ensemble.
Abstract
This paper describes the development of convolutional neural networks (CNN), a type of deep-learning method, to predict next-hour tornado occurrence. Predictors are a storm-centered radar image and a proximity sounding from the Rapid Refresh model. Radar images come from the Multiyear Reanalysis of Remotely Sensed Storms (MYRORSS) and Gridded NEXRAD WSR-88D Radar dataset (GridRad), both of which are multiradar composites. We train separate CNNs on MYRORSS and GridRad data, present an experiment to optimize the CNN settings, and evaluate the chosen CNNs on independent testing data. Both models achieve an area under the receiver-operating-characteristic curve (AUC) well above 0.9, which is considered to be excellent performance. The GridRad model achieves a critical success index (CSI) of 0.31, and the MYRORSS model achieves a CSI of 0.17. The difference is due primarily to event frequency (percentage of storms that are tornadic in the next hour), which is 3.52% for GridRad but only 0.24% for MYRORSS. The best CNN predictions (true positives and negatives) occur for strongly rotating tornadic supercells and weak nontornadic cells in mesoscale convective systems, respectively. The worst predictions (false positives and negatives) occur for strongly rotating nontornadic supercells and tornadic cells in quasi-linear convective systems, respectively. The performance of our CNNs is comparable to an operational machine-learning system for severe weather prediction, which suggests that they would be useful for real-time forecasting.
Abstract
This paper describes the development of convolutional neural networks (CNN), a type of deep-learning method, to predict next-hour tornado occurrence. Predictors are a storm-centered radar image and a proximity sounding from the Rapid Refresh model. Radar images come from the Multiyear Reanalysis of Remotely Sensed Storms (MYRORSS) and Gridded NEXRAD WSR-88D Radar dataset (GridRad), both of which are multiradar composites. We train separate CNNs on MYRORSS and GridRad data, present an experiment to optimize the CNN settings, and evaluate the chosen CNNs on independent testing data. Both models achieve an area under the receiver-operating-characteristic curve (AUC) well above 0.9, which is considered to be excellent performance. The GridRad model achieves a critical success index (CSI) of 0.31, and the MYRORSS model achieves a CSI of 0.17. The difference is due primarily to event frequency (percentage of storms that are tornadic in the next hour), which is 3.52% for GridRad but only 0.24% for MYRORSS. The best CNN predictions (true positives and negatives) occur for strongly rotating tornadic supercells and weak nontornadic cells in mesoscale convective systems, respectively. The worst predictions (false positives and negatives) occur for strongly rotating nontornadic supercells and tornadic cells in quasi-linear convective systems, respectively. The performance of our CNNs is comparable to an operational machine-learning system for severe weather prediction, which suggests that they would be useful for real-time forecasting.
