Introducing Deucalion and Pyrrha v1.0: Image Datasets for Disaster Management of Floods

Current paper, introduces Deucalion and Pyrrha, two image datasets for disaster management (DM) of floods. Deucalion v1.0 is consisted of 10240 photos. Main acquisition sources were Instagram, Kaggle datasets, video frames extracted from YouTube videos, Flickr photos and various search engine queries. The photos have been classified in two categories: I. Related to DM and II. Not related to DM. The manually classified photos were stored into two separate folders. Pyrrha v 1.0, is consisted of 2004 photos common to Deucalion, of both classes, further processed, extracting thus segmented features useful to DM. The features were accumulated in 20 different classes, described in the manuscript. Pyrrha v1.0 was consisted of 11,393 segmented and annotated features. A VGG-19, a ResNet101, and an EfficientNetB0 DL models were fine-tuned for binary image classification, using Deucalion v1.0. Moreover, specific classes of Pyrrha were used for training a YOLO v11 for image segmentation. The selected classes were “flooded” and “people”. The model parameters along with the corresponding training and validation accuracies, precisions and losses per epoch were visualized in related graphs and tables. Deucalion and Pyrrha is a set of the very few flood-related datasets, with analytic description, and a wide range of detail, considering the 20 classes of Pyrrha. Moreover it includes captures that can be found in either social media, news-videos, and fieldwork, in diverse flood disastrous events, around the globe. Validation precision values were above 0.95 in binary image classification, while object and mask detection in the flooded class received precision above 0.92. Deucalion and Pyrrha v1.0, are expected to be emerged as significant datasets specially in fields that require rapid extraction and dissemination.