資料內(nèi)容：
Highlights
Sub-meter resolution canopy height maps using self-supervised learn-
ing and a vision transformer trained on Aerial and GEDI Lidar
Jamie Tolan1, Hung-I Yang1, Ben Nosarzewski1, Guillaume Couairon2 , Huy
Vo2 , John Brandt3, Justine Spore3, Sayantan Majumdar4, Daniel Haziza2,
Janaki Vamaraju1, Theo Moutakani2, Piotr Bojanowski2, Tracy Johns1, Brian
White1, Tobias Tiecke1, Camille Couprie2
•
0.5 meter resolution canopy height maps at jurisdictional scale are re-
leased.
•
Improved performance from Self-Supervised Learning (SSL) and vision
transformers.
•
First use of SSL and vision transformers for canopy height estimation.
•
Low resolution GEDI and high resolution aerial lidar predictions are
combined.
•
Model generalises well to aerial imagery, even though trained with satel-
lite images.
