Self‐Supervised Classification of Weather Systems Based on Spatiotemporal Contrastive Learning

Abstract

Classification of weather systems (CWS) refers to the categorization of high-dimensional multivariate meteorological data into a reasonable and manageable number of typical weather systems that share similar meteorological fields, physical characteristics and evolutionary trends. Therefore, CWS has been broadly applied in weather forecasts and statistical climatology. In ensemble forecasts, CWS is utilized to generate a set of pre-defined circulation types and simplify the ensemble forecasts, where each weather pattern of ensemble members is assigned to the closest matching type, to a sequence of circulation type probabilities (

A Customized Multi‐Scale Deep Learning Framework for Storm Nowcasting

“…Additionally, our model shows poor performance on specific area where sample are much less than other areas (Section S8 in Supporting Information ). Self‐supervised learning (L. Wang et al., 2022) or transfer learning (Han et al., 2021) will be used for solving this kind of few‐shot issue. Critically, the work on loss functions highlights the importance of the “double penalty” in nowcasting.…”

Weather Systems Connecting Modes of Climate Variability to Regional Hydroclimate Extremes

“…While there have been many studies on WSC, there are two advantages in WAC‐hydro over similar models: (a) the predictand is the combination of P and T2, which is more hydrology relevant, compared to the phenomenological weather tags. This allows the model to be trained without manual labels and allows for explicit weather/hydrologic predictions (Gibson et al., 2016; Li et al., 2020; Xiao et al., 2021; Zhao et al., 2018); (b) existing quantitative WSC models usually focus on P alone, while our optimal number of weather anomaly modes are determined by optimizing the concurrent P‐T2 predictive skills, which enhances their hydrological applications (Kholodovsky & Liang, 2021; L. Wang et al., 2022; Wilson et al., 1991). For example, C1–C3 feature P‐driven floods, while C8–C9 feature floods resulting from P and rain‐on‐snow, which is sensitive to T2.…”

“…These types of clustering do not directly translate into hydrological applications, or tend to poorly characterize certain types of extreme events (Gibson et al., 2017). Meanwhile, clustering models targeted at land surface meteorological condition predictions are usually optimized to maximize predictive skills on a single variable, most often precipitation (Nishiyama et al., 2007; L. Wang et al., 2022; Xiao et al., 2021; Zhao et al., 2018). Such a sole focus on a single surface meteorological variable as predictand limits the consideration of other key meteorological variables for hydrologic predictions such as temperature (T2).…”

“…WSC has a long history and has been implemented in a manual way, automated way (e.g., correlation‐based or eigenvector‐based analysis), or a hybrid of both (Brinkmann, 2000; Esteban et al., 2006; Muller, 1977; Sheridan, 2002). With the rising popularity of deep learning and image recognition techniques, advanced models have been developed in the past decade to summarize the high‐dimensional weather conditions in reduced order space (Chattopadhyay et al., 2020; Elhoseiny et al., 2015; L. Wang et al., 2022; Zhang et al., 2018).…”

“…• A new weather system clustering model is developed specifically for hydroclimate analysis and prediction • Twelve weather systems in the Puget Sound region connect the regional hydroclimate to major modes of climate variability • Regional cold season floods are tied to two weather systems, each highlighting unique flood mechanisms (Gibson et al, 2017). Meanwhile, clustering models targeted at land surface meteorological condition predictions are usually optimized to maximize predictive skills on a single variable, most often precipitation (Nishiyama et al, 2007;L. Wang et al, 2022;Xiao et al, 2021;Zhao et al, 2018).…”