Research data

Machine Learning Datasets

*Updated: 22 of July 2026.
*Deucalion v1.0-anonymized can be delivered within 1 day or within a day from the time I will be at the storage location. 

*All the 10,240 photos of both classes are anonymized, by using the smurge tool of Gimp and a pen pad. That's the only difference from the Deucalion v1.0 dataset of the official publication (not the preprint).

*After cross checks minimal errors were found in Deucalion v1.0 dataset, which consist of less than 0.5% of the total. For reasons of consistency those have not been revised but annotated in a csv.  In case you want to correct them, you can do it easily by consolidating the csv and filtering the errors. In that case you will have Deucalion v1.0-anonymized-sp1.

*In case you find other mistakes, not indicated, then you will receive a 50$ refund per mistaken image. In that case for credibility reasons, you are requested not to correct the dataset alone in case you want to use it for official published research. You will have a future Deucalion v1.0-anonymized-sp2 version granted for free, whenever that version will be available.

*In case you want it for operational use though probably there is no credibility risk if you want to modify the dataset. But still it is not recommended.

*Discounts can be made from 10% and in rare cases up to 100%. Feel free to ask.

*Discounts can be made for developing countries as well and for Disaster Management Entities with proven low budget. Feel free to ask through mail.

**3 August 2026

**Pyrrha dataset v1.0-anonymized can be delivered within 1 day. You will receive: The 2004 photos, divided as train (1804 photos) and test size (200 photos) as per official paper. Contains also the 11.393 segments. The only difference from the official paper's data is that the photos have been anonymized.

It is delivered at a YOLO format with a ready to train structure for YOLO models of the ultralytics library, for the classes "flooded" and "people". 

Includes also the names of the rest of the classes and all the segments of all classes for further setup on your side.

In case you prefer other template than the YOLO-friendly one, mention it in the e-mail of the order. Depending on work load it can be delivered, free of charge within very few days.

Mapping Datasets

All the mapping datasets are 100% ready and can be sent to you as soon as I am in the office.