Fascinating work and inspiring application of the underlying DINOv3 image segmentation model!
The blog post and paper [1] describe a promising approach to solving related problems at previously impossible scale and quality: I am currently exploring methods to better represent seasonal land cover changes that would improve wind power generation forecasting and this paper provides a great starting point.
I hope DINOv3 can inspire more work like this - and I would encourage any curious mind to play with that model! I was amazed by its capability to distinguish between fine object details. For example, in a photo of a bicycle, the patch embeddings cleanly separated the background from the individual spokes of the wheel.
> CHMv2 is derived from single-date imagery, where the acquisition process selects the best available image within a target period (2017 -2020). This limits the direct use of the released CHMv2 data for attributing
canopy height to a specified year of interest. To support change applications, we provide the image acquisition date associated
with each prediction in the dataset metadata.
So generally a few years out of date, but the dataset is transparent about when each image was taken.
Related: Just the other day I used USGS 3DEP LiDAR data + Claude Code to get a sense for the number of trees on my property. Diffing terrain map and canopy map gives tree elevation. It was a fun project to explore, primarily because I set CC loose and said "here is the bounding box of my property, pad it by 50 feet and then go absolutely nuts against government datasets gathering as much open data as you can" - it figured out the rest. Dug into soil maps, historical satellite imagery, and lidar data.
I think they were buying carbon offsets at some point and trying to validate that the countries and organizations that were selling the carbon offset were not cutting down those trees, effectively profiting twice.
Presumably the smart ones just sell their promise-not-to-cut-down-my-forest multiple times. Laundered through completely trustworthy NGOs, so nothing can actually be audited properly.
The blog post and paper [1] describe a promising approach to solving related problems at previously impossible scale and quality: I am currently exploring methods to better represent seasonal land cover changes that would improve wind power generation forecasting and this paper provides a great starting point.
I hope DINOv3 can inspire more work like this - and I would encourage any curious mind to play with that model! I was amazed by its capability to distinguish between fine object details. For example, in a photo of a bicycle, the patch embeddings cleanly separated the background from the individual spokes of the wheel.
[1] https://arxiv.org/abs/2603.06382
> CHMv2 is derived from single-date imagery, where the acquisition process selects the best available image within a target period (2017 -2020). This limits the direct use of the released CHMv2 data for attributing canopy height to a specified year of interest. To support change applications, we provide the image acquisition date associated with each prediction in the dataset metadata.
So generally a few years out of date, but the dataset is transparent about when each image was taken.
The tree outside of house is not 9 feet tall per. I have a 2 story house and it easily towers 10 feet higher than my house.
Additionally, there are several Royal Palms that are close to 50ft and they show as being only 15 feet.
Here are the visuals re: trees - https://i.imgur.com/R0W4q4O.png
What did you do to actually count trees? Even from aerial Lidar it can be a bit finicky for closed canopies.