BIG data methods for improving windstorm FOOTprint prediction (BigFoot)

点击翻译按钮获取中文摘要

Wind storms can cause great damage to property and infrastructure. The windstorm footprint (a map of maximum wind gust speed over 3 days) is an important summary of the hazard of great relevance to the insurance industry and to infrastructure providers. Windstorm footprints are conventionally estimated from meteorological data and numerical weather model analyses. However there are several interesting less structured data sources that could contribute to the estimation of the wind storm footprints, and more importantly will raise the spatial resolution of our estimates. This is important as there are important small-scale meteorological phenomena, such as sting jets, that are currently not well resolved by the current methods. We propose to exploit three additional sources of data (and possibly others during the course of the project). The three sources so far identified identified are amateur observations available through the Met Office weather observations website (WOW), comments made on social media and video recorded on social media or CCTV. Amateur meteorological observations are currently collected by the Met Office but not used in producing the footprint estimates. We will investigate whether we can use them in the estimation of the storm footprint; a useful by-product will be estimates of the uncertainty for each WOW station. Social media, such as twitter or instagram, often contains comments on windstorms. These can range from comments on how windy it is, to reports of damage produced by storms. In some cases the geographical location of the message is provided by the device but in others it has to be inferred. There are very large numbers of messages posted on social media every day and it should be possible to used these to provide more detailed modelling of footprints. In addition to text, social media also records images and video. Video is also recorded extensively in the form of CCTV. Video recordings of trees, say, blowing in the wind include information on the strength of the windstorm. We will analyse such recordings to produce information on wind velocity and gust velocity. Bringing together large quantities of diverse data is a complex procedure. We will develop, test, and compare two approaches in modern data science: statistical process modelling and machine learning. Both methods will aim to synthesise all the data into an estimate of the windstorm footprint (and its associated uncertainty). The former will concentrate on producing a map more like the current estimates based on the maximum gust speed while the latter data based methods will concentrate more on mapping the damage caused by the storm. Once we have estimates of the windstorm footprint from both social media and the modelling we will compare these with the standard products and, in consultation with stakeholder, establish any improvements.