The increasing amount of falling rain may cause several problems especially in urban areas, which drainage system can often not handle this large amount in a short time. Confirming a flooded scene in a timely manner can help the authorities to take further actions to counter the crisis event or to get prepared for future relevant incidents. This paper studies the detection of flood events comparing two successive in time Sentinel-2 images, a method that can be extended for detecting floods in a time-series. For the flood detection, fine-tuned pre-trained Deep Convolutional Neural Networks are used, testing as input different sets of three water sensitive satellite bands. The proposed approach is evaluated against different change detection baseline methods, based on remote sensing. Experiments showed that the proposed method with the augmentation technique applied, improved significantly the performance of the neural network, resulting to an F-Score of 62% compared to 22% of the traditional remote sensing techniques. The proposed method supports the crisis management authority to better estimate and evaluate the flood impact.