In recent years, social media outlets have been widely exploited for disaster analysis and retrieving relevant information. Social media information can help in several ways, such as finding the mostly affected areas and information on casualties and scope of the damage etc. In this
paper, we tackle a specific facet of social media in natural disasters, namely the identification of passable routs in a flooded region. In detail, we propose several solutions for two relevant tasks, namely (i) identification of flooded and non-flooded images in a collection of images retrieved
from social media, and (ii) identification of passable roads in a flooded region. To this aim, we mainly rely on existing deep models pre-trained on ImageNet and Places dataset, where the models pre-trained on ImageNet extract object specific and the ones pre-trained on places dataset extract
scene-level features. In order to properly utilize the object and scene-level features, we rely on different fusion methods including Particle Swarm Optimization (PSO) and Genetic Modeling of the deep features in a late fusion manner. The evaluation of the proposed methods are carried out
on the large-scale datasets provided for MediaEval- 2018 benchmarking competition on Multimedia and Satellites. The results demonstrate significant improvement in the performance over the baselines.