We present a novel unsupervised Deep Learning method for flood segmentation in Unmanned Aerial Vehicle imagery. This method utilizes automatically generated labels as masks for the training process, eliminating the need for actual ground truth data. On a public dataset consisting of 290 RGB images, we train two well-known Convolutional Neural Network architectures, typically used for semantic segmentation, to perform flood segmentation. The proposed method is tested and compared in a totally unknown public dataset consisting of 663 RGB images, yielding high-performance results, with 92.2% and 88.2% overall accuracy and F1-score, respectively.