Africa Biomass Challenge
In the graph below I present the result of the model that got me to the 7th spot in the leaderboard.
- In blue the actual AGBDs from the test set (sorted by increasing values)
- In orange are my predictions
- In green the mean AGBD from the test set
- In red the mean AGBD from my predictions
As we can see, the orange line is slightly better than predicting the mean (the model predict slightly lower than mean AGBD for very low AGBD images and slighly higher than mean AGBD for very high AGBD images), but this is barely better than average and very far from the actual AGBD.
What do you guys think ? Do you find your models to have the same behavior : basically predicting the mean ? Or can they actually identify increasing AGBD from the images ?
[EDIT] Oops, I thoug I could upload images in the discussion. Here is a link : https://postimg.cc/mzdTdxQS
Very nice insight. However, I could not understand one thing. Is the test set labels (AGBD) also available ?
So, the actual test set AGBD are not available, this would ruin the purpose of the competition. The ones I shown are from an unseen during training split took from the training dataset, sorry if this wasn't clear.
I think we have the same issue. It is not really adapting to the curve of true biomass