It may be possible, though it seems quite unlikely to get rmse of 8 before preprocessing.
The approach of removing part of 2020 is novel - many of the images in that subset were "initial" photos of empty ground which all had 0 extent. However, they look similar to damaged crop fields, which often had very high extent values.
I attempted to model this and saw small improvements; removing the images altogether could yield further gains
He has already explained his approach in this discussion
https://zindi.africa/competitions/cgiar-eyes-on-the-ground-challenge/discussions/19794
If you cheated, you can draw a dreamlike solution :)
It may be possible, though it seems quite unlikely to get rmse of 8 before preprocessing.
The approach of removing part of 2020 is novel - many of the images in that subset were "initial" photos of empty ground which all had 0 extent. However, they look similar to damaged crop fields, which often had very high extent values.
I attempted to model this and saw small improvements; removing the images altogether could yield further gains