Improving the credibility of unreliable information through static images and data mining: an experimental study to identify floods

Affected locations by flash floods are rich in information for flood management. Usually, there are several types and sources of information which can be related to achieve better reliability to decision-making. However, a major challenge is to achieve the reliability this information within datasets so heterogeneous or complex. For instance, reports of Volunteered Geographic Information (VGI) through a crowdsourcing-based platform can be confirmed by means of images available on site. Thus, we carried out an experiment to identify water level of river through clustering from static images using an evolutionary method of hierarchical data clustering, called DAta-MIning COde REpositories 1 (Sanches, Cardoso, and Delbem, 2011). Our experiment aimed answering the following question: Is DAMICORE able to find matching clusters between static images gathered from the sensor in-situ and water levels provided by the non-automatic interpretation mechanisms in the riverbed? These mechanisms (Figure 1 (a) water level ruler, (b) puppet, and (c) multi-color band) refer to the hazard index at hydrology field and help volunteers to report into the crowdsourcing-based platform (Degrossi, Albuquerque, Fava, Mendiondo, 2014). Moreover, our dataset 2 contains 288 images categorized in (gray) 124 undefined, (orange) 109 acceptable, (red) 17 high, (dark red) 6 very high, and (blue) 32 flood which were obtained from insitu sensor in a 5-minute temporal resolution on November 23rd, 2015, when a flash flood occurred at 4 p.m. (Figure 1 (d) ) at São Carlos, São Paulo, Brazil. Our preliminaries results have shown a possible matching between clusters found (Figure 2) and interpretation mechanisms of the water level in the riverbed. Therefore, there is evidence that the DAMICORE can support VGI reports collected from dedicated platforms, improving the credibility of information. Nevertheless, further experiments should be performed considering a greater number of images per category and matching between other types of VGI and authoritative data, e.g. social media and sensor.