events in the mid-latitudes. We use convolutional neural networks (CNN) to draw
warm and cold fronts in gridded weather data. CNN’s are a method from deep
learning, a subset of machine learning where the model builds representations of the
input data at various levels of abstraction. CNN’s are specially designed to handle
spatiotemporal grids, running convolutional filters over a grid at different resolutions
to detect features at different scales.
The predictors used are 3-D wind vectors, density, and wet-bulb potential temper-
ature from the North American Regional Reanalysis. The verification data (“ground
truth”) consist of human-drawn fronts from the Weather Prediction Center (WPC).
We have conducted experiments to determine the best predictor variables, pre-
processing methods, and model configurations.
The output of the CNN consists of three probabilities at each grid cell: no front
(NF), warm front (WF), and cold front (CF). However, these grids are hard to
interpret (humans usually think of fronts as objects rather than probability grids) and
often contain overly wide frontal zones (regions of elevated WF or CF probability).
This is alleviated by converting the probability grids to objects. Specifically, we
convert regions of high WF and CF probability to polygons, find the main skeleton
line of each polygon, and then smooth the skeleton line to make it resemble a human-
drawn front.
Forecast quality is evaluated in two settings. The first is pixelwise, where pixel
(i, j) in the forecast grid is compared to pixel (i, j) in the verification grid. The
verification fronts are dilated with a 50-km buffer, so this corresponds roughly to
neighbourhood evaluation with a 50-km neighbourhood distance, which does not
unduly penalize the model for small spatial or temporal errors. The best model
achieves a Peirce score of 0.704, and Gerrity score of 0.700, on the testing data.
Both scores range from -1. . .1, where 1 indicates a perfect model and 0 indicates no
skill. However, the frequency bias of this model is ~16, which means that the forecast
data contain ~16 times as many frontal pixels as the verification data.
The second forecast-evaluation setting is object-based. Each forecast WF (CF)
polyline is matched to the nearest WF (CF) in the verification data, as long as the
two are within some neighbourhood distance. This reduces the frequency bias to ~1
(unbiased).
We have extended previous work in frontal detection by (a) assessing predictor
importance, (b) combining deep learning with object-based methods, and (c) per-
forming detailed forecast evaluation on a large dataset. Our approach could be used
in a research setting – e.g., to create a long-term climatology of fronts, including their
spatiotemporal and morphological characteristics – and in operational forecasting,
to quickly visualize and compare fronts produced by different numerical models.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner