The genesis module from our global tropical cyclone model was released earlier this month, and provides a good opportunity to showcase some of the building blocks behind our modelling philosophy. Although the model is global, the discussion here will focus on the North Atlantic Hurricane component only and on a single one of the climate indicator that influences the forecast. Let’s call it feature A.
Feature engineering
As for many of our climate indicators, feature A was constructed using a neural network based compression algorithm called the autoencoder. This allows efficient representation of the observed variability in selected global atmospheric fields, and for feature A the focus is on seasonal anomalies of the global precipitation field since 1970 (i.e. the post-satellite era).
Some of the key patterns identified by feature A are summarised below where the yellow (blue) coloured cells show strong positive (negative) weights. In other words, a strong positive anomaly in seasonal precipitation for cells coloured in yellow (blue) contribute to an increase (decrease) in the value of feature A.
Fig. 1: Sign (yellow: positive / blue: negative) of strongest weights involved in the autoencoder layer responsible for Feature A
Our automated predictor selection engine identifies the value of feature A during the March-May season (FA-MAM) as one of the strongest indicators of hurricane activity during the following June-November season.
Fig. 2: relationship between the March-May value of feature A and the number of named hurricanes during the following June-November season
Any physics?
Since the construction and selection of feature A as a key driver of Hurricane activity both result from a fully automated process, no a-priori “expert knowledge” about the physics of Hurricane formation is explicitly included. Instead we let the data speak (i.e. a machine learning approach). Yet testing for consistency between the relationship selected and our understanding of what physical mechanisms contribute most to Hurricane formation is a very valuable (and comforting) exercise. Here is what we found:
- With large feature A values being associated with above average precipitation around Indonesia and below average precipitation in the eastern Pacific (Fig. 1) it is safe to describe feature A as a measure of the ENSO signal: seasons with a large feature A value are very likely to be associated with la Niña events (a well-accepted driver of increased hurricane activity).
- At a more fundamental level we also investigated the correlation between FA-MAM and anomaly fields of environmental variables known to influence Hurricane activity in August-October (ASO: peak of hurricane season). FA-MAM values correlate strongly with warmer than average ASO sea surface temperatures (Fig. 3a), reduced wind shear conditions (Fig. 3b), and lower than normal sea level pressures (Fig. 3c) in the North Atlantic tropical basin: all widely accepted as factors conducive to enhanced hurricane activity.
Fig.3: Coefficient of correlation between the May value of feature A (FA March-May) and August-October anomalies of (a) sea surface temperature, (b) zonal wind shear and (c) sea level pressure.
- The anomaly patterns on exhibit in Fig. 3 also suggest FA-MAM values can provide an early indication of the strength of the ENSO signal in ASO: large positive FA-MAM values correlate with warmer than average waters, enhanced wind shear and a low-pressure environment around Indonesia; they also tend to precede a cooler than average South East Pacific Ocean within a high-pressure environment in ASO – all indicators of a strong potential for the presence of a la Niña event.
How is all this useful?
At the very least some early indication of whether to expect an above, within or below average Hurricane activity in the June-November season can be gained from the value of feature A at the end of the MAM season. In particular a high hurricane activity is more likely in years where the MAM anomaly of precipitation in regions such as the North of Brazil or Indonesia is high (i.e. areas covered by yellow cells in Fig. 1) or if precipitation is anomalously low in the Eastern Pacific, North East Brazil or in the southern parts of Africa (blue colored cells in Fig. 1).
At reask we rely on a family of models called Bayesian hierarchical models to help us quantify what “more likely” means and handle the high level of uncertainty that characterise seasonal forecasting. Using inputs from a whole variety of climate indicators such as feature A (covering a range of fields, atmospheric levels and lead times) we are able to compute the probability of certain levels of Hurricane risk during the June-November season, conditional on the observed state of the climate in earlier months.
If this article was of interest please stay tuned for our 2018 Northern Hemisphere TC forecast to be released early June.