Search

Over the past decade the use of machine learning in meteorology has grown rapidly. Specifically neural networks and deep learning have been used at an unprecedented rate. To fill the dearth of resources covering neural networks with a meteorological lens, this paper discusses machine learning methods in a plain language format that is targeted to the operational meteorological community. This is the second paper in a pair that aim to serve as a machine learning resource for meteorologists. While the first paper focused on traditional machine learning methods (e.g., random forest), here a broad spectrum of neural networks and deep learning methods is discussed. Specifically, this paper covers perceptrons, artificial neural networks, convolutional neural networks, and U-networks. Like the Part I paper, this manuscript discusses the terms associated with neural networks and their training. Then the manuscript provides some intuition behind every method and concludes by showing each method used in a meteorological example of diagnosing thunderstorms from satellite images (e.g., lightning flashes). This paper is accompanied with an open-source code repository to allow readers to explore neural networks using either the dataset provided (which is used in the paper) or as a template for alternate datasets. [ABSTRACT FROM AUTHOR]

Providing timely warnings for severe and potentially tornadic convection is a critical component of the NWS mission, and owing to the associated large reflectivity gradients, sidelobe contamination is possible. This paper focuses on elevation sidelobe contamination appearing in the low-level inflow region of supercells. A qualitative conceptual model of the Weather Surveillance Radar-1988 Doppler (WSR-88D) antenna pattern interacting with supercells is introduced, along with Doppler power spectrum representations of the potential mix of returned power from the main lobe and the sidelobes. These tools inform the multiple ways elevation sidelobe contamination appears in the low levels, primarily below 3 km (10 kft) of radar data. The most common manifestation is somewhat noisy data similar to particulates or biota in clear air. Trained NWS forecasters are accustomed to mentally filtering out noisy clear-air returns as less important. Elevation sidelobe contamination can be mixed with the three-body scatter spike (TBSS) artifact, though the TBSS remains the more salient feature. The most consequential form is the apparent circulation, and when it is incorrectly interpreted as valid, contributes to the false alarm ratio (FAR) for NWS tornado warnings. Quantitative results on the effect of elevation sidelobe contamination on FAR are presented. Diagnostic techniques are emphasized, and with familiarization, can be used in real-time warning operations to identify the apparent circulation as either valid or an imposter. Identification of these contaminated velocity signatures offers a unique opportunity to reduce the NWS tornado warning FAR without also reducing the probability of detection (POD). Significance Statement: The WSR-88D weather radars provide overall high-quality data for users. However, with some severe thunderstorms, an artifact called elevation sidelobe contamination can produce what looks like a rotation signature, but it may not be real. These ambiguous velocity signatures can contribute to tornado warnings based on rotation signatures that are false circulations. This paper specifically focuses on elevation sidelobe contamination due to its impact on tornado warning decisions. Diagnostic techniques, including several examples, are presented here to aid the reader in correctly identifying elevation sidelobe contamination and why it may occur. Correct identification of an apparent circulation as an imposter due to contamination is a unique opportunity to improve NWS tornado warning performance by reducing warning false alarms. [ABSTRACT FROM AUTHOR]

A time-space shift method is developed for relocating model-predicted tornado vortices to radar-observed locations to improve the model initial conditions and subsequent predictions of tornadoes. The method consists of the following three steps. (i) Use the vortex center location estimated from radar observations to sample the best ensemble member from tornado-resolving ensemble predictions. Here, the best member is defined in terms of the predicted vortex center track that has a closest point, say at the time of t 5 t*, to the estimated vortex center at the initial time t0 (when the tornado vortex signature is first detected in radar observations). (ii) Create a time-shifted field from the best ensemble member in which the field within a circular area of about 10-km radius around the vortex center is taken from t 5 t*, while the field outside this circular area is transformed smoothly via temporal interpolation to the best ensemble member at t0. (iii) Create a time-space-shifted field in which the above time-shifted circular area is further shifted horizontally to co-center with the estimated vortex center at t0, while the field outside this circular area is transformed smoothly via spatial interpolation to the non-shifted field at t0 from the best ensemble member. The method is applied to the 20 May 2013 Oklahoma Newcastle-Moore tornado case, and is shown to be very effective in improving the tornado track and intensity predictions. SIGNIFICANCE STATEMENT: The time-space shift method developed in this paper can smoothly relocate tornado vortices in model-predicted fields to match radar-observed locations. The method is found to be very effective in improving not only model initial condition but also the subsequent tornado track and intensity predictions. The method is also not sensitive to small errors in radar-estimated vortex center location at the initial time. The method should be useful for future real-time or even operational applications although further tests and improvements are needed (and are planned). [ABSTRACT FROM AUTHOR]

This paper proposes a new quantitative precipitation estimation (QPE) technique to provide accurate rainfall estimates in complex terrain, where conventional QPE has limitations. The operational radar QPE in Taiwan is mainly based on the simplified relationship between radar reflectivity Z and rain rate R [R(Z) relation] and only utilizes the single-point lowest available echo to estimate rain rates, leading to low accuracy in complex terrain. Here, we conduct QPE using deep learning that extracts features from 3D radar reflectivities to address the above issues. Convolutional neural networks (CNN) are used to analyze contoured frequency by altitude diagrams (CFADs) to generate the QPE. CNN models are trained on existing rain gauges in northern and eastern Taiwan with the 3-yr data during 2015–17 and validated and tested using 2018 data. The weights of heavy rains (≥10 mm h−1) are increased in the model loss calculation to handle the unbalanced rainfall data and improve accuracy. Results show that the CNN outperforms the R(Z) relation based on the 2018 rain gauge data. Furthermore, this research proposes methods to conduct 2D gridded QPE at every pixel by blending estimates from various trained CNN models. Verification based on independent rain gauges shows that the CNN QPE solves the underestimation of the R(Z) relation in mountainous areas. Case studies are presented to visualize the results, showing that the CNN QPE generates better small-scale rainfall features and more accurate precipitation information. This deep learning QPE technique may be helpful for the disaster prevention of small-scale flash floods in complex terrain. [ABSTRACT FROM AUTHOR]

While storm identification and tracking algorithms are used both operationally and in research, there exists no single standard technique to objectively determine performance of such algorithms. Thus, a comparative skill score is developed herein that consists of four parameters, three of which constitute the quantification of storm attributes—size consistency, linearity of tracks, and mean track duration—and the fourth that correlates performance to an optimal postevent reanalysis. The skill score is a cumulative sum of each of the parameters normalized from zero to one among the compared algorithms, such that a maximum skill score of four can be obtained. The skill score is intended to favor algorithms that are efficient at severe storm detection, i.e., high-scoring algorithms should detect storms that have higher current or future severe threat and minimize detection of weaker, short-lived storms with low severe potential. The skill score is shown to be capable of successfully ranking a large number of algorithms, both between varying settings within the same base algorithm and between distinct base algorithms. Through a comparison with manually created user datasets, high-scoring algorithms are verified to match well with hand analyses, demonstrating appropriate calibration of skill score parameters. Significance Statement: With the growing number of options for storm identification and tracking techniques, it is necessary to devise an objective approach to quantify performance of different techniques. This study introduces a comparative skill score that assesses size consistency, linearity of tracks, mean track duration, and correlation to an optimal postevent reanalysis to rank diverse algorithms. This paper will show the capability of the skill score at highlighting algorithms that are efficient at detecting storms with higher severe potential, as well as those that closely resemble human-perceived storms through a comparison with manually created user datasets. The novel methodology will be useful in improving systems that rely on such algorithms, for both operational and research purposes focusing on severe storm detection. [ABSTRACT FROM AUTHOR]

The analysis and forecast of precipitation characteristics is a key task for national meteorological services for providing high-quality weather forecasts and warnings. Beside the precipitation amount, the precipitation type is essential to describe and evaluate the recent, ongoing, and future weather situations. This paper introduces a new surface-based hybrid hydrometeor classification algorithm. The presented method combines polarimetric radar observations at radar beam height from the C-band dual-polarization weather radar network of the German Weather Service [Deutscher Wetterdienst (DWD)] with corrected thermodynamical profiles of numerical weather prediction (NWP) model output and extrapolates the hydrometeor classes at radar beam height to a height of 2 m above ground level (AGL). The implemented technique parameterizes the microphysical processes in the lower troposphere based on the appropriate thermodynamical profile that the hydrometeors have to pass along their way from the radar beam height to the surface. Due to errors in NWP output, the NWP vertical profiles of temperature and humidity are adjusted by using several types of surface stations with high spatial and temporal resolution. Verification results show considerable improvements in the hydrometeor classification near the ground compared to the radar sweep height. After an additional positive in-house evaluation, the presented method is integrated into DWD's operational environment. The topic of this paper is to describe the processing steps for the computation of the near-surface precipitation type. In addition, example cases and a verification study complement the explanations. Significance Statement: For a national weather service and its customers, the separation of precipitation types into solid and liquid is a key issue for improving weather forecasts and warnings. We combined measurements from weather radars and weather stations together with output of numerical weather prediction models and developed a method of estimating the precipitation type at a height of 2 m. Previous studies of the hydrometeor classification field concentrate on warm season weather conditions using "radar observations only" or apply a hybrid combination of radar observations with model information. The uniqueness of our work is the additional introduction of surface observations with high temporal and spatial resolution for a better estimation of the thermodynamical profiles to improve the "all season alignment." [ABSTRACT FROM AUTHOR]

The highly successful fleet of Weather Surveillance Radar-1988 Doppler (WSR-88D) radars is approaching its end of service, and research efforts are under way to inform a decision toward a possible WSR-88D replacement. A methodology to link radar design characteristics to impacts on how radar data are used to diagnose hazardous weather was developed through a unique partnership between radar-engineering innovations in radar simulations and National Weather Service (NWS) radar data interpretation expertise. Deep commitment to two-way learning across disciplines resulted in a methodology that is both efficient and highly relevant to the NWS hazardous weather warning program. The methodology presented in this paper is a model for revealing complex trade-offs between weather radar characteristics and their resultant impact on NWS data interpretation for threat identification. This qualitative methodology is presented in the context of a broader proof-of-concept study from which it was developed. Adapted for further research, it can support the crucial role of deriving quantitative radar design criteria that balance the trade-offs among radar capabilities, cost, and impact to users. That is, the proposed methodology supports the evaluation of candidates for a potential WSR-88D replacement and any necessary major system upgrades in the interim. Significance Statement: Defining the requirements for an operational weather radar system is ideally achieved with clearly identified trade-offs among cost, radar design characteristics, and impacts on user data interpretation. This work is an advancement of the historic evolution of weather radar development to support the U.S. National Weather Service (NWS) mission, based on collaboration among researchers, radar engineers, and NWS forecasters. The methodology presented here is an adaptable tool for revealing these essential, but complex trade-offs, providing a roadmap for further studies toward the next-generation NWS weather radar fleet. [ABSTRACT FROM AUTHOR]

Infrequent lightning flashes occurring outside of surface precipitation pose challenges to Impact-Based Decision Support Services (IDSS) for outdoor activities. This paper examines the remote sensing observations from an event on 20 August 2019 where multiple cloud-to-ground flashes occurred over 10 km outside surface precipitation (lowest radar tilt reflectivity, 10 dBZ and no evidence of surface precipitation) in a trailing stratiform region of a mesoscale convective system. The goal is to demonstrate the fusion of radar with multiple lightning observations and a lightning risk model to demonstrate how reflectivity and differential reflectivity combined provided the best indicator for the potential of lightning where all of the other lightning safety methods failed. A total of 13 lightning flashes were observed by the Geostationary Lightning Mapper (GLM) within the trailing stratiform region between 2100 and 2300UTC. The average size of the 13 lightning flashes was 3184 km², with an average total optical energy of 7734 fJ. A total of 75 NLDN flash locations were coincident with the 13GLMflashes, resulting in an average of 5.8 NLDN flashes [in-cloud (IC) and cloud-to-ground (CG)] per GLM flash. In total, five of the GLM flashes contained at least one positive cloud-to-ground flash (1CG) flash identified by the NLDN, with peak amplitudes ranging between 66 and 136 kA. All eight CG flashes identified by the NLDN were located more than 10 km outside surface precipitation. The only indication of the potential of these infrequently large flashes was the presence of depolarization streaks in differential reflectivity (ZDR) and enhanced reflectivity near the melting layer. [ABSTRACT FROM AUTHOR]

Machine learning algorithms are trained on a 10-yr archive of composite weather radar images in the Swiss Alps to nowcast precipitation growth and decay in the next few hours in moving coordinates (Lagrangian frame). The hypothesis of this study is that growth and decay is more predictable in mountainous regions, which represent a potential source of practical predictability by machine learning methods. In this paper, artificial neural networks (ANN) are employed to learn the complex nonlinear dependence relating the growth and decay to the input predictors, which are geographical location, mesoscale motion vectors, freezing level height, and time of the day. The average long-term growth and decay patterns are effectively reproduced by the ANN, which allows exploring their climatology for any combination of predictors. Due to the low intrinsic predictability of growth and decay, its prediction in real time is more challenging, but is substantially improved when adding persistence information to the predictors, more precisely the growth and decay and precipitation intensity in the immediate past. The improvement is considerable in mountainous regions, where, depending on flow direction, the root-mean-square error of ANN predictions can be 20%–30% lower compared with persistence. Because large uncertainty is associated with precipitation forecasting, deterministic machine learning predictions should be coupled with a model for the predictive uncertainty. Therefore, we consider a probabilistic perspective by estimating prediction intervals based on a combination of quantile decision trees and ANNs. The probabilistic framework is an attempt to address the problem of conditional bias, which often characterizes deterministic machine learning predictions obtained by error minimization. [ABSTRACT FROM AUTHOR]

This paper investigates the ability of the Weather Research and Forecasting (WRF) Model in simulating multiple small-scale precipitation bands (multibands) within the extratropical cyclone comma head using four winter storm cases from 2014 to 2017. Using the model output, some physical processes are explored to investigate band prediction. A 40-member WRF ensemble was constructed down to 2-km grid spacing over the Northeast United States using different physics, stochastic physics perturbations, different initial/boundary conditions from the first five perturbed members of the Global Forecast System (GFS) Ensemble Reforecast (GEFSR), and a stochastic kinetic energy backscatter scheme (SKEBS). It was found that 2-km grid spacing is adequate to resolve most snowbands. A feature-based verification is applied to hourly WRF reflectivity fields from each ensemble member and the WSR-88D radar reflectivity at 2-km height above sea level. The Method for Object-Based Diagnostic Evaluation (MODE) tool is used for identifying multibands, which are defined as two or more bands that are 5–20 km in width and that also exhibit a >2:1 aspect ratio. The WRF underpredicts the number of multibands and has a slight eastward position bias. There is no significant difference in frontogenetical forcing, vertical stability, moisture, and vertical shear between the banded versus nonbanded members. Underpredicted band members tend to have slightly stronger frontogenesis than observed, which may be consolidating the bands, but overall there is no clear linkage in ambient condition errors and band errors, thus leaving the source for the band underprediction motivation for future work. [ABSTRACT FROM AUTHOR]

The local, linear, least squares derivative (LLSD) approach to radar analysis is a method of quantifying gradients in radar data by fitting a least squares plane to a neighborhood of range bins and finding its slope. When applied to radial velocity fields, for example, LLSD yields part of the azimuthal (rotational) and radial (divergent) components of horizontal shear, which, under certain geometric assumptions, estimate one-half of the two-dimensional vertical vorticity and horizontal divergence equations, respectively. Recent advances in computational capacity as well as increased usage of LLSD products by the meteorological community have motivated an overhaul of the LLSD methodology's application to radar data. This paper documents the mathematical foundation of the updated LLSD approach, including a complete derivation of its equation set, discussion of its limitations, and considerations for other types of implementation. In addition, updated azimuthal shear calculations are validated against theoretical vorticity using simulated circulations. Applications to nontraditional radar data and new applications to nonvelocity radar data including reflectivity at horizontal polarization, spectrum width, and polarimetric moments are also explored. These LLSD gradient calculations may be leveraged to identify and interrogate a wide variety of severe weather phenomena, either directly by operational forecasters or indirectly as part of future automated algorithms. [ABSTRACT FROM AUTHOR]

Raindrop size sorting is a ubiquitous microphysical occurrence in precipitating systems. Owing to the greater terminal fall speed of larger particles, a raindrop's fall trajectory can be sensitive to its size, and strong air currents (e.g., a convective updraft) can enhance this sensitivity. Indeed, observational and numerical model simulation studies have confirmed these effects on raindrop size distributions near convective updrafts. One striking example is the lofting of liquid drops and partially frozen hydrometeors above the environmental 0°C level, resulting in a small columnar region of positive differential reflectivity ZDR in polarimetric radar data, known as the ZDR column. This signature can serve as a proxy for updraft location and strength, offering operational forecasters a tool for monitoring convective trends. Beneath the 0°C level, where WSR-88D spatiotemporal resolution is highest, anomalously high ZDR collocated with lower reflectivity factor at horizontal polarization ZH is often observed within and beneath convective updrafts. Here, size sorting creates a deficit in small drops, while relatively large drops and melting hydrometeors are present in low concentrations. As such, this unique raindrop size distribution and its related polarimetric signature can indicate updraft location sooner and more frequently than the detection of a ZDR column. This paper introduces a novel algorithm that capitalizes on the improved radar coverage at lower levels and automates the detection of this size sorting signature. At the algorithm core, unique ZH–ZDR relationships are created for each radar elevation scan, and positive ZDR outliers (often indicative of size sorting) are identified. Algorithm design, examples, performance, strengths and limitations, and future development are discussed. [ABSTRACT FROM AUTHOR]

Thirty National Weather Service forecasters worked with 1-, 2-, and 5-min phased-array radar (PAR) volumetric updates for a variety of weather events during the 2015 Phased Array Radar Innovative Sensing Experiment. Exposure to each of these temporal resolutions during simulated warning operations meant that these forecasters could provide valuable feedback on how rapidly updating PAR data impacted their warning decision processes. To capture this feedback, forecasters participated in one of six focus groups. A series of open-ended questions guided focus group discussions, and forecasters were encouraged to share their experiences and opinions from the experiment. Transcriptions of focus group discussions were thematically analyzed, and themes belonging to one of two groups were identified: 1) forecasters' use of rapidly updating PAR data during the experiment and 2) how forecasters envision rapidly updating PAR data being integrated into warning operations. Findings from this thematic analysis are presented in this paper, and to illustrate these findings from the forecasters' perspectives, dialogue that captures the essence of their discussions is shared. The identified themes provide motivation to integrate rapidly updating radar data into warning operations and highlight important factors that need to be addressed for the successful integration of these data. [ABSTRACT FROM AUTHOR]

The 2013 Phased Array Radar Innovative Sensing Experiment (PARISE) investigated the impacts of higher-temporal-resolution radar data on National Weather Service forecasters' warning decision processes during severe hail and wind events. In total, 12 forecasters participated in the 2013 PARISE over a 6-week period during the summer of 2013. Participants were assigned to either a control [5-min phased-array radar (PAR) updates] or experimental (1-min PAR updates) group, and worked two cases in simulated real time. This paper focuses on the qualitative retrospective reports of participants' warning decision processes that were collected using the recent case walk-through method. Timelines of participants' warning decision process were created for both cases, which were then thematically coded according to a situational awareness framework. Coded themes included perception, comprehension, and projection. It was found that the experimental group perceived significantly more information during both cases than the control group (case 1 p = 0.045 and case 2 p = 0.041), which may have improved the quality of their comprehensions and projections. Analysis of timelines reveals that 1-min PAR updates were important to the experimental group's more timely and accurate warning decisions. Not only did the 1-min PAR updates enable experimental participants to perceive precursor signatures earlier than control participants, but through monitoring trends in radar data, the experimental group was able to better detect storm motion, more accurately identify expected weather threats from severe thunderstorms, more easily observe strengthening and diminishing trends in storms, and make more correct tornado-related warning decisions. [ABSTRACT FROM AUTHOR]

Observations and recent high-resolution numerical model simulations indicate that liquid water and partially frozen hydrometeors can be lofted considerably above the environmental 0°C level in the updrafts of convective storms owing to the warm thermal perturbation from latent heating within the updraft and to the noninstantaneous nature of drop freezing. Consequently, upward extensions of positive differential reflectivity (i.e., ZDR ≥ 1 dB)-called ZDR columns-may be a useful proxy for detecting the initiation of new convective storms and examining the evolution of convective storm updrafts. High-resolution numerical simulations with spectral bin microphysics and a polarimetric forward operator reveal a strong spatial association between updrafts and ZDR columns and show the utility of examining the structure and evolution of ZDR columns for assessing updraft evolution. This paper introduces an automated ZDR column algorithm designed to provide additional diagnostic and prognostic information pertinent to convective storm nowcasting. Although suboptimal vertical resolution above the 0°C level and limitations imposed by commonly used scanning strategies in the operational WSR-88D network can complicate ZDR column detection, examples provided herein show that the algorithm can provide operational and research-focused meteorologists with valuable information about the evolution of convective storms. [ABSTRACT FROM AUTHOR]

The Geostationary Operational Environmental Satellite-14 ( GOES-14) Imager operated in 1-min Super Rapid Scan Operations for GOES-R (SRSOR) mode during summer and fall of 2012 to emulate the high temporal resolution sampling of the GOES-R Advanced Baseline Imager (ABI). The current GOES operational scan interval is 15-30 min, which is too coarse to capture details important for severe convective storm forecasting including 1) when indicators of a severe storm such as rapid cloud-top cooling, overshooting tops, and above-anvil cirrus plumes first appear; 2) how satellite-observed cloud tops truly evolve over time; and 3) how satellite cloud-top observations compare with radar and lightning observations at high temporal resolution. In this paper, SRSOR data, radar, and lightning observations are used to analyze five convective storms, four of which were severe, to address these uncertainties. GOES cloud-top cooling, increased lightning flash rates, and peak precipitation echo tops often preceded severe weather, signaling rapid intensification of the storm updraft. Near the time of several severe hail or damaging wind events, GOES cloud-top temperatures and radar echo tops were warming rapidly, which indicated variability in the storm updraft that could have allowed the hail and wind gusts to reach the surface. Above-anvil cirrus plumes were another prominent indicator of impending severe weather. Detailed analysis of storms throughout the 2012 SRSOR period indicates that 57% of the plume-producing storms were severe and 85% of plumes from severe storms appeared before a severe weather report with an average lead time of 18 min, 9 min earlier than what would be observed by GOES operational scanning. [ABSTRACT FROM AUTHOR]

This paper documents features that led to major forecast errors on the 12-24-h time scale in the nature and location of severe weather in the southern plains on 30 May 2012. Evidence is presented that the forecast errors were the result of 1) dry air that originated in a region of dissipating, elevated convective storms, and which was advected in a narrow tongue into western Oklahoma, inhibiting convective initiation; 2) the development of a cyclone along the dryline in western Texas, to the east of which several supercells formed; 3) the upscale development of the supercells into a mesoscale convective system (MCS) at nightfall; and 4) the dissipation of an MCS that had formed along a cold front in southwestern Kansas and was propagating into northwestern Oklahoma, as it encountered dry, subsiding air underneath the stratiform precipitation region of the rear portion of the MCS farther south. There was a meridionally oriented swath of high winds in clear air, in between the two MCSs. This swath of high winds may have been associated with a bore triggered at night by the MCSs approaching from the north, as the MCS collapsed, producing a gust front that propagated through stable, low-level air. This case study illustrates how the predictability of severe weather in a region can be extremely sensitive to the details of where nearby convective storms form and how they evolve. It also highlights the likely importance of the accurate representation of cloud microphysics and dynamics in numerical forecast models on predictability. [ABSTRACT FROM AUTHOR]