CHMV2: Mapping the World’s Forest Canopies at 1-Meter Resolution

We explore how the World Resources Institute and Meta built CHMV2, a global canopy height map at 1-meter resolution. Using a self-supervised AI that predicts 3D depth from 2D satellite imagery, anchored by independent tree detection and aligned with airborne laser scans, this project overcomes seasonal and platform misalignment to reveal fine-scale forest structure and empower restoration, agroforestry, and biodiversity efforts.

Note:  This podcast was AI-generated, and sometimes AI can make mistakes.  Please double-check any critical information.

Sponsored by Embersilk LLC

Transcript

SPEAKER_01 0:00

Last weekend, um, I spent an embarrassing amount of time arguing with my neighbor. Oh no. Yeah. It was over the exact height of this oak tree that basically straddles our property line. I mean, I was out there with a tape measure trying to use like high school trigonometry on the shadow it was casting.

SPEAKER_00 0:17

Aaron Powell Well, nature doesn't exactly come with a ruler attached.

SPEAKER_01 0:20

No, it does not. And he was using this smartphone app that just kept crashing on him.

SPEAKER_00 0:24

Right.

SPEAKER_01 0:24

So we eventually just gave up. But it really highlighted something for me, which is that measuring just one single tree is surprisingly difficult.

SPEAKER_00 0:31

It really is. I mean, we like precision, but a tree is a highly complex living structure.

SPEAKER_01 0:38

Aaron Powell Exactly. Which brings us to our mission for this deep dive. Because if measuring one backyard oak is that hard, I mean, imagine trying to measure literally every tree on Earth.

SPEAKER_00 0:48

Aaron Powell That's a massive, massive scale.

SPEAKER_01 0:50

Aaron Powell Right. And we are looking at this truly awe-inspiring collaboration between the World Resources Institute and Meta. They've built something called CHM V2, which is uh a one-meter resolution map of the entire Globe's Forest Canopy using advanced AI.

SPEAKER_00 1:06

Aaron Powell Yeah. And to really grasp what a leap forward this is, you have to understand what we were working with before.

SPEAKER_01 1:12

Which was pretty blurry, right?

SPEAKER_00 1:13

Oh, very. Previous global canopy maps had a resolution of maybe, you know, 10 to 30 meters per pixel. They essentially acted like an old pixelated video game. Just big chunky blocks. Exactly. Totally missing short vegetation or the nuanced structure of a forest. So CHM V2 changes the game entirely by using this advanced self-supervised AI model called DNO V3 to basically predict 3D depth from flat, two-dimensional MAXR satellite imagery.

SPEAKER_01 1:43

Aaron Powell Wait, self-supervised? Does that mean there isn't like a team of interns sitting in a basement manually labeling millions of tree photos to teach the AI what to look for?

SPEAKER_00 1:51

Aaron Powell Pretty much, yeah. The AI just feeds on massive amounts of raw data and learns the underlying visual patterns on its own.

SPEAKER_01 1:57

Aaron Powell So it's looking at a flat 2D satellite photo. Is it basically learning to squint at the picture? Like um, you know those old magic eye posters from the 90s? Oh, yeah. Where if you stare at a flat pattern long enough, the 3D structure just magically pops out?

SPEAKER_00 2:09

Aaron Powell The Magic Eye is a really good analogy, actually. But you know, it's doing something much more mathematical. It is essentially reverse engineering the angle of the sun based on how shadows fall across the canopy. Oh, wow. Yeah. It reads the texture and the way light hits the leaves to calculate physical height. But to teach the AI how to do that math, you have to give it an answer key first. You train it on perfectly matched 3D laser data, which is known as airborne laser scanning or ALS.

SPEAKER_01 2:38

Okay, wait, I'm stuck here. Because these satellites and airplanes, they aren't exactly flying in tight formation.

SPEAKER_00 2:44

No, not at all.

SPEAKER_01 2:45

Right. So if the satellite takes a 2D picture in the summer when the canopy is like full and green, and then the laser scans the 3D data in the winter when the branches are bare, how can the AI possibly align those two totally different images?

SPEAKER_00 2:58

Well, that is the exact nightmare the developers faced. It is a massive alignment puzzle. The data is often taken months or even years apart.

SPEAKER_01 3:05

Yeah, that sounds impossible.

SPEAKER_00 3:07

If you feed that messy, mismatched data to the AI, it gets totally confused and just spits out a blurry prediction.

SPEAKER_01 3:13

So how do you anchor them together without, I don't know, manually adjusting millions of photos?

SPEAKER_00 3:18

Aaron Ross Powell They got really clever. They use an independent AI tree detection model to draw bounding boxes around individual trees. Oh, okay. Yeah. And those boxes act as geographical anchor points. So they pull the disparate 2D optical data sets and the 3D laser data sets into this beautifully unified grid.

SPEAKER_01 3:36

Aaron Powell So they basically built an AI matchmaker to force misaligned puzzle pieces together.

SPEAKER_00 3:42

That's a perfect way to put it, yeah.

SPEAKER_01 3:43

You know, aligning massive, messy data sets from totally different time periods is a huge headache. And actually, it's the exact kind of data integration problem that our sponsor, Embersilk, helps companies solve.

SPEAKER_00 3:54

Oh, nice.

SPEAKER_01 3:55

Yeah. If you need help with AI training or automation or integration or software development, they are the experts. Uncovering where agents could make the most impact for your business or personal life. Check out Embersilk.com for AI needs. So, anyway, once they finally use those bounding boxes to lock the data into place, what happens next?

SPEAKER_00 4:14

Well, that is when the model's true potential is just unleashed. The results are stunning. I mean, CHMV2 dramatically reduces errors for the giants of the forest. You know, the trees over 30 meters tall.

SPEAKER_01 4:26

The really big ones.

SPEAKER_00 4:27

Exactly. Right. And it flawlessly captures fine-scale structures. We're talking about mapping individual canopy gaps and edges with literally one meter precision.

SPEAKER_01 4:37

Which means we can actually monitor forest restoration, optimize agroforestry, and just support global biodiversity on a micro level everywhere.

SPEAKER_00 4:46

It really shows what happens when global collaboration meets machine learning. We are taking raw, chaotic data and turning it into a precise structural understanding of our planet. It allows us to appreciate the beauty of Earth's ecosystems like never before.

SPEAKER_01 5:01

That is just so incredibly optimistic. If we can map the exact height of complex forest canopies from space right now, I mean, imagine the discoveries waiting for you when we can track the daily growth of a single newly planted sapling anywhere on Earth.

SPEAKER_00 5:14

That's an amazing thought.

SPEAKER_01 5:15

It really is. I might never have to use trigonometry on my neighbor's oak tree again.

SPEAKER_00 5:19

Let's hope not.

SPEAKER_01 5:20

If you enjoyed this podcast, please subscribe to the show. Hey, leave us a five star review if you can. It really does help get the word out. Thanks for tuning in.