Automated Front Detection - Using computer vision and machine learning to explore a new direction in automated weather forecasting
Not peer reviewed
MetadataShow full item record
In weather forecasting, automation and computing are the driving forces of innovation. More computing power and better techniques allow for faster and more accurate weather data systems. The task of detecting fronts (interfaces between different air masses) in weather systems has yet to be solved computationally with such accuracy. In computer science and information science research, the techniques in artificial intelligence used for pattern recognition are constantly evolving and solving new problems, both in the weather domain and elsewhere. I therefore explore whether artificial intelligence can be used to help detecting fronts in weather systems, as well as what weather features are useful to study in this endeavor. In my Master’s project I have developed an automatic front detection system in cooperation with weather service provider StormGeo, under the Design Science Research paradigm. The study aims to further our understanding of AI techniques and their use in weather analysis, through the design, development and use of an information system. The research follows in the footsteps of recent developments in several research fields, both within meteorology, weather prediction, data modelling, computer vision and machine learning. The system development was based on core principles of agile and lean software development methodologies, and used commonly available tools and techniques. The resulting system identifies fronts using computer vision techniques, and classifies them using machine learning techniques and expert knowledge in meteorology. The system is fairly accurate in finding the major front lines in a weather system, and is even able to find some fronts that meteorologists have missed, but it fails to pick up many subtle details that expert use in front detection. The system excels at classifying some types of fronts, but performs poorly on others. Geopotential height, air temperature, specific humidity and relative vorticity are the weather features used by the system, that most accurately predicts the location of fronts, although other features could be used successfully as well. This project could outline a new, computer driven way of discovering fronts in weather data, based on known concepts from computer vision. However, the techniques are in need of more development and refinement to be able to compete with expert human analysis, and to be employed in full scale by the industry. These developments and refinements should, however, be achievable with today’s technology, given adequate time and resources. Finally, the project raises the discussion of the need of an objective, absolute definition of fronts, based on common front indicators, to objectively and quantitatively evaluate and further improve front detection systems of all types.
PublisherThe University of Bergen
Copyright the Author. All rights reserved