Development of a Human–Machine Mix for Forecasting Severe Convective Events

Abstract

Providing advance warning for impending severe convective weather events (i.e., tornadoes, hail, wind) fundamentally requires an ability to predict and/or detect these hazards and subsequently communicate their potential threat in real time. The National Weather Service (NWS) provides advance warning for severe convective weather through the issuance of tornado and severe thunderstorm warnings, a system that has remained relatively unchanged for approximately the past 65 years. Forecasting a Continuum of Environmental Threats (FACETs) proposes a reinvention of this system, transitioning from a deterministic product-centric paradigm to one based on probabilistic hazard information (PHI) for hazardous weather events. Four years of iterative development and rapid prototyping in the National Oceanic and Atmospheric Administration (NOAA) Hazardous Weather Testbed (HWT) with NWS forecasters and partners has yielded insights into this new paradigm by discovering efficient ways to generate, inform, and utilize a continuous flow of information through the development of a human–machine mix. Forecasters conditionally used automated object-based guidance within four levels of automation to issue deterministic products containing PHI. Forecasters accomplished this task in a timely manner while focusing on communication and conveying forecast confidence, elements considered necessary by emergency managers. Observed annual increases in the usage of first-guess probabilistic guidance by forecasters were related to improvements made to the prototyped software, guidance, and techniques. However, increasing usage of automation requires improvements in guidance, data integration, and data visualization to garner trust more effectively. Additional opportunities exist to address limitations in procedures for motion derivation and geospatial mapping of subjective probability.

Nowcasting lightning occurrence from commonly available meteorological parameters using machine learning techniques

“…Lagerquist et al 42 described a machine learning system that forecasts the probability of damaging straight-line wind for each storm cell in the continental United States. Karstens et al 44 developed a human-machine mix for forecasting severe convective events. As an application in lightning nowcasting and early warning systems, this paper examines how the mining of basic atmospheric datasets can be used to explore correlation patterns between lightning incidence and atmospheric data and, thus, for nowcasting of lightning activity.…”

Techniques and Thresholds of Significance for Using WSR-88D Velocity Data to Anticipate Significant Tornadoes

“…Some promise of methods for determining whether a storm will continue to produce significant damage (a key consideration of IBWs), also was shown. Future work may be able to better refine this relationship, particularly with regard to "Warn-on-Forecast" type operations under development within NOAA and the National Severe Storms Laboratory (Karstens et al 2018;Rothfusz et al 2018). The results also show that QLCS/MCV significant tornadoes continue to present a significant predictability challenge and that additional effort in this area may help improve overall NWS tornado warning performance.…”

“…This seems to validate existing guidance provided by the NWS WDTD that uses a so-called "30-40-50" rule for first guess IBW decision making (WDTD 2018b). This training emphasizes peak skill scores from S15, T17, G16, and Kingfield and LaDue (2015), indicating that V rot ≥25.72 m s -1 (50 kt) is a threshold that should be concerning for the potential existence or imminent threat of a significant supercell tornado, regardless of other cues. Note that this is the V rot at peak RSC anywhere in the storm (excluding storm top divergence signals, which were not considered part of the mesocyclone; WDTD 2018a), and not just the lowest available elevation angle.…”

“…Toth et al (2013) published initial research suggesting that such discrimination was possible. Kingfield and LaDue (2015) were among the first to specifically focus on IBW execution, and determined that automated mesocyclone detections provided the highest skill scores in discriminating between weak (EF0 and EF1) and significant (EF2+) tornadoes at rotational velocity [V rot = (V max -V min )/2), where V max is the maximum outbound wind speed and V min is the most negative (i.e., inbound) wind speed] values between 18 and 25.72 m s -1 (35 and 50 kt). There was considerable overlap between weak and significant tornado events, but overall the skill scores were comparable to the skill in differentiating between tornadic and nontornadic signatures.…”

Forecasting Severe Weather with Random Forests

“…In addition to severe weather, ML has demonstrated success in forecasting heavy precipitation (e.g., Gagne et al 2014;Herman and Schumacher 2018a,b;Whan and Schmeits 2018;Loken et al 2019), cloud ceiling and visibility (e.g., Herman and Schumacher 2016;Verlinden and Bright 2017), and tropical cyclones (Loridan et al 2017;Alessandrini et al 2018;Wimmers et al 2019). Furthermore, automated probabilistic guidance, including ML algorithms, have been identified as a priority area for integrating with the operational forecast pipeline (e.g., Rothfusz et al 2014;Karstens et al 2018). However, many past applications have focused on either much shorter time scales, such as nowcast settings (e.g., Marzban and Stumpf 1996;Lagerquist et al 2017), or on much longer time scales (e.g., Tippett et al 2012;Elsner and Widen 2014;Baggett et al 2018), with less emphasis on the day-ahead time frame and very little model development in the medium-range (e.g., Alvarez 2014).…”