I'd like to start by thanking you for organizing this competition on this really interesting theme.
After starting to get encouraging results, I'm starting to have some problems with the data. I am going to explain them in 4 parts: training labels, data redundancy, leak between training and validation sets and overfitting. In order to properly illustrate my points, I have written a short detailed document with images. Here is the link to read it:
https://docs.google.com/document/d/e/2PACX-1vTXKjBSY_eeCzi79HyDT64hzQkLucg851DjT7kS5m_pV5eWLprnVhrXi8_bQTLyy2wXxGotezIioGVH/pub
To conclude, I'd like to know if other participants are experiencing the same kind of problems and get some feedback on this short, rapid analysis.
Thank you in advance!
Thank you for your message and the clarity of your feedback.
We are in the process of investigating. The fact is that all the inconsistencies you report are the result of simulated flooded areas. These are not always visible on the images, as the flooded areas may be more or less masked by vegetation.
We are currently checking the simulations in greater detail, and will keep you informed as soon as possible.