Automated detection of weather fronts using a deep learning neural network

Abstract. Deep learning (DL) methods were used to develop an algorithm to

automatically detect weather fronts in fields of atmospheric surface

variables. An algorithm (DL-FRONT) for the automatic detection of fronts was

developed by training a two-dimensional convolutional neural network (2-D CNN)

with 5 years (2003–2007) of manually analyzed fronts and surface fields

of five atmospheric variables: temperature, specific humidity, mean sea

level pressure, and the two components of the wind vector. An analysis of

the period 2008–2015 indicates that DL-FRONT detects nearly 90 % of the

manually analyzed fronts over North America and adjacent coastal ocean

areas. An analysis of fronts associated with extreme precipitation events

shows that the detection rate may be substantially higher for important

weather-producing fronts. Since DL-FRONT was trained on a North American

dataset, its extensibility to other parts of the globe has not been tested,

but the basic frontal structure of extratropical cyclones has been applied

to global daily weather maps for decades. On that basis, we expect that

DL-FRONT will detect most fronts, and certainly most fronts with significant

weather. However, where complex terrain plays a role in frontal orientation

or other characteristics, it might be less successful.