Twitter use helps predict flooding
Heavy rainfall can cause streets to flood and basements and tunnels to overflow. Jan van Rijn investigated, together with Christiaan Lamers (formerly of Leiden University) and Ton Beenen (STOWA, RIONED), how data science can help to predict which areas are at greater risk of flooding. Van Rijn presented the first results on 28 January at the congress of the umbrella organization for urban water management, RIONED.
Using the KNMI precipitation radar measurements, the General Altitude Map of the Netherlands and reports of rainwater nuisance on Twitter, they developed a model that can make predictions about water nuisance. ‘By combining the precipitation measurements with the terrain characteristics, you know all the factors that occur during flooding,’ explains Van Rijn. ‘But in order for a computer model to predict where flooding will actually occur, it needs examples. We therefore used Tweets in which flooding is discussed. Because there is a GPS signal attached to the Tweets, we can combine these messages with the data we get from the maps and the model can learn which data lead to Tweets about flooding and which do not. Eventually, this should allow the model to predict, for a given location and amount of precipitation, whether flooding will occur.'
Points for improvement
The first results seem positive. Van Rijn shows that the accuracy of the model is 57.9%: ‘If the model were to guess, it should be 50%, because there are only two possibilities: either flooding or no flooding. That means that the model sees connections between the data and that this way of working has potential.’ Van Rijn does emphasize, however, that the model is still a concept and will have to be developed further: ‘This is really only a first step.’
To be able to use the model in practice, adjustments are needed. ‘Tweets are a fairly unreliable way of measuring where flooding occurs,’ explains Van Rijn. ‘To be able to make accurate and good predictions, we will have to use other data. Think, for example, of damage claims received by insurance compagnies, satellite data or even sensors in the city that measure how much water is on the streets at given times. Another point for improvement is that in this version we worked with the total amount of precipitation that falls on a day, while it obviously makes a lot of difference to the occurrence of flooding whether that precipitation is spread over the entire day or falls in one hour. We therefore want to include this information in a subsequent version so that the forecasts become more accurate.’
To prevent is better than to cure
Although the model is still mainly a concept, Van Rijn sees possibilities: ‘This model shows the potential of data science in water management. With a computer model, we will soon be able to predict which places are vulnerable, so that we can focus more on prevention. For example, we can carry out predictive maintenance to sewers and dikes or solve problems in the drainage of water. Moreover, it can give us an insight into the effects of drought, which we have also had to deal with more frequently in recent years.’