How do you reconcile these two statements, because it seems that there will not be a flood within 40-60 weeks on either end of the 730 days??
...events contain randomly select 40-60 weeks prior and post flood date, to ensure the flood event date is not always the middle index....
and
Somewhere in these 730 days per location/event ID there could be a flood.... If there is a flood for that event it can happen on any ONE of those 730 days.
The training data contains indeed at most one single flood day more or less in the middle of the 730 days period for each event. Maybe they meant that the testing data could have that flood day at any given day. Anyhow I don't see this as a problem, just don't bias your model into predicting a flood in the middle of the period