The governments of the Myanmar, Lao PDR and Cambodia seek tools to overcome financing gaps that arises after disasters happen. For this purpose, The World Bank together with insurers are developing tooling to support parametric insurance (and bridge that gap), but there are omissions in the current data availability for such purpose. Therefore in this project we analyse the feasibility of online media to complement the loss estimation process and deliver a pilot modelling platform to test basic operational functionalities.
In the forthcoming months we will first review the existing social and news media data and its applicability to complement flood estimates in the countries of interest, i.e. Cambodia, Lao PDR and Myanmar. The coverage of social media and which platforms people are using differs per country. For instance Philippines and Indonesia have abundant Twitter coverage, while in Tanzania people are using especially Facebook, WhatsApp and a local forum called Jamiiforums. Besides social media, we will make inventory of other online media available such as news stations. During floods a large number of news items are shared online, often with great detail.
Next we start analysing the media data (while in large quantities, normally >100.000 documents per country). FloodTags designed a number of processes to extract event information from large datasets of media data, using a mix of artificial intelligence, natural language processing and combinations with external data sources, including satellite imagery. These processes will be adjusted and added to fit the loss-estimation process within this assignment. To cope with the resulting uncertainty we determine the reliability and relevance of the online media analysis by: 1) Assessing the confidence of the NLP classifications 2) Internal comparison of independent observations within a social network (do independent observers confirm a detected event) and 3) External comparison of the events with other data (for instance are detected events accompanied with extreme hydro-meteorological conditions or is the event located in flood prone areas).
Finally, together with the client we will assess the potential of using the online media data for the actual ground-truthing of flood events in the selected countries, for the purpose of parametric insurance. First results are expected medio 2018.