Content area
Full text
Machine learning model interpretation and visualization focusing on meteorological domains are introduced and analyzed.
Machine learning (ML) and deep learning (DL; LeCun et al. 2015) have recently achieved breakthroughs across a variety of fields, including the world’s best Go player (Silver et al. 2016, 2017), medical diagnosis (Rakhlin et al. 2018), and galaxy classification (Dieleman et al. 2015). Simple forms of ML (e.g., linear regression) have been used in meteorology since at least the 1950s (Malone 1955), and ML has been used extensively to forecast convective hazards since the mid-1990s. Kitzmiller et al. (1995) use linear regression to forecast the probability of tornadoes, large hail, or damaging wind; Billet et al. (1997) use linear regression to forecast hail probability and size; Marzban and Stumpf (1996, 1998) use neural networks to forecast the probability of tornadoes and damaging wind, respectively; and Marzban and Witt (2001) use neural networks to forecast hail size. Gagne et al. (2013, 2017a) use random forests to forecast hail probability at 1-day lead time; McGovern et al. (2014) and Williams (2014) use random forests to forecast convectively induced aircraft turbulence; while Cintineo et al. (2014, 2018) use naïve Bayes to forecast the probability of tornadoes, large hail, and damaging wind. DL is also beginning to be used in meteorology, with applications including hail prediction (Gagne et al. 2019) and detection of extreme weather patterns such as tropical cyclones, atmospheric rivers, and synoptic-scale fronts (Liu et al. 2016; Mahesh et al. 2018; Kunkel et al. 2018; Lagerquist et al. 2019b). The authors have extensive experience using ML to improve forecasting and understanding of weather phenomena (Gagne et al. 2017a,b; Lagerquist et al. 2017; McGovern et al. 2017; Gagne et al. 2019; Lagerquist et al. 2018). Many of these products have been used by human meteorologists in experiments and day-to-day operations.
Despite its wide adoption in meteorology, ML is often criticized by forecasters and other end users as being a “black box” because of the perceived inability to understand how ML makes its predictions. This phenomenon is not exclusive to meteorology, and many ML practitioners and users have recently begun to focus on this interpretability problem (Olah et al. 2017; Lipton 2016; NeurIPS Foundation 2018; Molnar 2018).
The main contribution of this paper is...