Yes definitely. Most importantly, the data must be public. If people post something to their friends alone, we cannot use it. If they post it on a public Facebook page from your organisation, we can ingest with little problems.

You are right that Twitter is not as popular in one place as it is in the other. However, we find that in regions with low Twitter counts, people still want to express themselves one way or the other. Besides a few tweets (we have found that there are always some tweets when it floods), there public messages are shared on blogs, forums, public Facebook pages, flickr etc. These sources can be ingested and made available for you. Do check out the blog we wrote about Tanzania, where we connected the software to online news media and to a popular (local) forum called Jamiiforums. 

Note that this is discussing the real-time media data, which comes in on very short timescales (within minutes after the start of an event). On a different timescale (within hours after start of an event) much more data becomes available from various news media. These describe the local situation often in great detail.

Indeed online media data is written by people, which makes the data source subjective. To still be able to use it as a trustworthy datasource, together with our scientific partners, we determine the reliability and relevance of the online media content in three steps

• By natural language processing (within the message we determine what is says)
• By comparing the messages against each other (different independent messages can confirm each other) and
• By comparing the messages with other data (messages after rain in low lying areas are often more relevant, than messages shared during a drought on high plains).

We have the precision of the natural language processing alone (without any combinations of data) now fluctuating between 70%-90% dependent of the source (this means that 10%-30% of messages we think are about floods, are actually about something else than floods) with an acceptable recall (number of false negatives, i.e. there are floods but they are not identified by us as floods). With various techniques and data combinations, precision can be increased to near 100%, with the same number of false negatives (some small events may always be missed). Depending on the use-case, we can optimize precision versus recall. For instance Philippine Red Cross has a high tolerance for false positives, but they are much averse to false negatives. Another Client (a museum we work for) has high tolerance for false negatives, but what is shown has to be true (no false positives).

Starting point in The Philippines was that (even without requesting people to report on a new floods) there are a massive amount of photos and observations circulating online. Rather that asking communities to report, we wanted to benefit from what is already out there. And this is the same approach we take in every new implementation: First let’s see what is out there that we can use!

Of course the next step is to engage with the people and the communities directly, and collect even more information. You can specifically request your volunteers or citizens actively to post about the situation they are in and the results would mean an even better understanding of the local situation. This is also what we are preparing now. Building on the information that we receive by passive listening, as FloodTags we deliver the data to:

• Target the regions where information is needed most: Regions that have a high risk but very little data coming in (for some regions we get tens of thousands tweets, perhaps not a priority to get even more data from there)
• Target the champions in those regions, that are likely to contribute. These can be users that already shared useful information, or users that have shown their public engagement otherwise on those regions.

Totally agree. Actually we need logical thinking people to “teach” the computer which articles (and what parts of it) they find important. If you think your staff could help in this that would be great! Via an interface we deliver (a website), they would need to point out which word is an event is interesting. E.g. the word “flood” between the word “new” and “hits” (making this “new flood hits”) indicates a new flood event, the number “3.000” before “evacuated” means that 3.000 people were evacuated. Etc. And this can become very specific for your organisations’ interest (what do YOU want to know). If your staff could help here that would be fantastic.

Secondly, for the easiness of the uptake, we would like to explore together with the staff on how to show the results of the software. E.g. we can show it in our own Dashboard, but also it could be connected to already running systems in your office (do you already have some kind of dashboard or systems). Similarly. the results of the impact forecasting (what impact could be expected at a given forecast) could be shared through a lookup table, which would make it very non-technical and easy to use without any further server or software requirements. To know the best way to share the results, we request the input of your staff.

The FloodTags platform is a hosted software solution that can be connected to via an API or front-end dashboard. Our revenue model comprises of

• Hours x tariff for new feature development and configuration of the platform and
• Subscriptions for the hosting and maintenance of (parts of) the platform

The code itself is shared with our Clients and can freely be used (so you can also choose to host and maintain the resulting software and configuration yourself). We are working on the license to go fully open source.