Assessing flood severity from georeferenced photos

Abstract

The use of georeferenced social media data in disaster and crisis management is increasing rapidly. Particularly in connection to flooding events, georeferenced images shared by citizens can provide situational awareness to emergency responders, as well as assistance to financial loss assessment, giving information that would otherwise be very hard to collect through conventional sensors or remote sensing products. Moreover, recent advances in computer vision and deep learning can perhaps support the automated analysis of these data. In this paper, focusing on ground-level images taken by humans during flooding events, we evaluate the use of deep convolutional neural networks for (i) discriminating images showing direct evidence of a flood, and (ii) estimating the severity of the flooding event. Considering distinct datasets (i.e., the European Flood 2013 dataset, and data from different editions of the MediaEval Multimedia Satellite Task), we specifically evaluated models based on the DenseNet and EfficientNet neural network architectures, concluding that these models for image classification can achieve a very high accuracy on both tasks.