CHMv2: Improvements in Global Canopy Height Mapping using DINOv3

The paper introduces CHMv2, a new global meter-resolution canopy height map that significantly improves accuracy and structural detail over existing products by leveraging a DINOv3-based depth estimation model trained on expanded, diverse airborne laser scanning data and validated against millions of satellite observations.

The Gist

Imagine you are trying to measure the height of every single tree in the entire world.

In the past, scientists had two main ways to do this:

1. The "Drone Survey": Sending planes equipped with super-accurate laser scanners (LiDAR) over forests. This is like sending a drone to measure every tree in your backyard with a ruler. It's incredibly precise, but it's expensive and slow. We only have these "drone maps" for a few lucky places (mostly the US, Europe, and parts of Asia).
2. The "Satellite Guess": Looking at the world from space using regular cameras. This covers the whole globe, but it's like trying to guess the height of a tree by looking at its shadow from a plane. It's often blurry, misses small details, and tends to underestimate how tall the really big trees are.

Enter CHMv2: The "Super-Translator"

This paper introduces CHMv2, a new global map that combines the best of both worlds. Think of it as a brilliant translator that learns to speak "Drone" (precise laser data) so it can understand "Satellite" (regular photos) and translate them into a 3D height map for the whole planet.

Here is how they built this new map, explained with some everyday analogies:

1. The Brain Upgrade: From "Smart" to "Genius"

The old map (CHMv1) used a smart AI brain called DINOv2. The new map uses DINOv3, which is like upgrading from a very smart high school student to a PhD candidate who has read every book in the library.

• The Analogy: Imagine teaching a child to recognize a dog. The old model needed to see 100 pictures of Golden Retrievers to understand what a dog looks like. The new model (DINOv3) has seen millions of unlabeled images of everything—cats, cars, trees, clouds—and has learned the essence of shapes and structures on its own. This allows it to recognize a tree's shape even if it's never seen that specific forest before.

2. The Training Camp: Cleaning the Mess

To teach the AI, the researchers needed to show it pairs of "Satellite Photo" and "Laser Height Map." But the data was messy.

• The Problem: Sometimes the photo was taken in summer (leaves on), and the laser map was taken in winter (leaves off). Sometimes the photo was slightly shifted to the left compared to the laser map. It's like trying to learn to bake a cake by looking at a photo of a cake that is slightly blurry and shifted to the side of the recipe book.
• The Fix: The team built an automated "cleaning crew." They used a smart detector to find the trees in the photos, find the trees in the laser maps, and then physically "slide" the laser map until the trees perfectly matched the photos. They also threw out the bad examples (like photos with clouds or mismatched seasons).
• The Result: The AI is now learning from a perfectly aligned, high-quality textbook rather than a messy pile of notes.

3. The Teacher's Strategy: A New Curriculum

The way they taught the AI changed, too.

• Old Way: The AI was told, "Just get the general shape right." This meant it was okay with being a little blurry or guessing the height of a giant tree as "medium."
• New Way: They introduced a special "curriculum."
  • Phase 1: First, they taught the AI to understand the relative heights (which tree is taller than the other) using a specific math trick.
  • Phase 2: Then, they switched to a different math trick that forced the AI to be precise about the actual numbers, especially for very tall trees.
  • Phase 3: They added a "sharpness" test. If the AI drew a tree edge that looked fuzzy, they gave it a penalty. This forced the map to have crisp, sharp edges, just like a high-definition photo.

Why Does This Matter? (The "So What?")

1. It sees the details.
Previous global maps were like looking at a forest from a helicopter; you could see the green blob, but not the gaps between trees or the edges of the canopy. CHMv2 is like standing on a hill with binoculars. It can see individual tree crowns, gaps where trees fell, and the jagged edges of a forest. This is crucial for spotting illegal logging or monitoring how well a forest is recovering.

2. It stops underestimating the giants.
Old maps often said a 60-meter tall tree was only 40 meters tall. This matters because taller trees hold more carbon. CHMv2 is much better at measuring these "giants," giving us a more accurate count of how much carbon the world's forests are storing.

3. It works everywhere.
Because the AI learned from a diverse mix of forests (from the Amazon to the US, from plantations to wild jungles), it doesn't get confused when it sees a new type of forest. It's no longer biased toward just American or European trees.

The Bottom Line

CHMv2 is a 1-meter resolution map of the entire world's tree heights.

Think of it as the difference between a pixelated, low-res sketch of a forest and a 4K, high-definition 3D model. It helps scientists, governments, and conservationists answer big questions: How much carbon is this forest holding? Is this agroforestry system healthy? Where is the forest degrading?

It's a massive leap forward in our ability to "see" the health of our planet's lungs, all thanks to a smarter AI, cleaner data, and a better way of teaching it.

Technical Summary

1. Problem Statement

Accurate, high-fidelity canopy height maps are critical for quantifying forest carbon, monitoring restoration/degradation, and assessing habitat structure. However, existing global canopy height products face three primary limitations:

• Low Resolution: Many products are medium-resolution (10m–30m), failing to capture fine-scale structural metrics (e.g., canopy edges, gaps, heterogeneity) essential for detailed forest management.
• Geographic Bias: Models trained on Airborne Laser Scanning (ALS) data often inherit strong geographic priors because ALS coverage is uneven (concentrated in North America, Europe, and Asia), leading to poor performance in underrepresented forest types.
• Data Misalignment: Fusing optical imagery with ALS-derived Canopy Height Models (CHMs) is challenging due to temporal mismatches (phenology changes) and spatial misregistration (geolocation errors), which degrade the learning of fine structural details.

Previous work (CHMv1) made strides but still suffered from underestimation of tall forests, blurring of canopy edges, and residual biases.

2. Methodology

The authors present CHMv2, a global, 1-meter resolution canopy height map derived from high-resolution optical satellite imagery (Maxar Vivid2) using a depth-estimation model built on the DINOv3 self-supervised learning (SSL) backbone.

A. Data Curation and Registration

• Training Data Expansion: The training corpus was significantly expanded and diversified, moving beyond the limited NEON (USA) data used in CHMv1. It now includes:
  • NAIP-3DEP: ~300k high-quality pairs from the US (NAIP optical + 3DEP ALS).
  • SatLidar v2: A curated, registered revision of a global dataset (~726k train samples) covering diverse geographies.
  • NAIP Sea: ~3,500 samples of open water/shorelines to prevent false height predictions over water.
• Automated Cleaning: A few-shot anomaly detector using DINOv3 embeddings was employed to filter out poor-quality pairs (e.g., land-use mismatches, sensor artifacts), removing ~15% of the data.
• Advanced Registration: To address spatial misalignment between optical and ALS data, the authors developed a two-step registration pipeline:
  1. Local Alignment: Uses tree detection bounding boxes (from a DINO DETR model) as control points to compute dense warp fields, correcting local shifts.
  2. Global Alignment: Uses FFT-based cross-correlation on high-canopy peak masks to correct rigid 2D translations.
  • Result: 85% of training samples with non-zero canopy height benefited from alignment, significantly reducing label noise.

B. Model Architecture and Training

• Backbone: Replaced the DINOv2-H encoder with the more capable DINOv3 Sat-L backbone, leveraging improved semantic representations from large-scale unlabeled imagery.
• Decoder Architecture: Modified the decoder to handle 448×448 input resolution (up from 256×256), introduced mixed linear/log binning strategies, and increased hidden layer dimensions.
• Curriculum Loss Function: A novel loss formulation was designed to address the specific distribution of canopy heights (many zeros, sparse tall structures):
  1. SiLog Loss: Used initially to establish relative depth structure.
  2. Charbonnier Loss: Gradually introduced to correct high-depth bias and encourage linear space accuracy.
  3. Patch Gradient Loss: A multi-scale gradient loss (pixel magnitude, patch range, and direction consistency) added mid-training to enhance structural sharpness without introducing grid artifacts.
• Sampling Strategy: Implemented category-based sampling to ensure batches contain a balanced mix of short (<1m) and tall (>25m/35m) trees, preventing the model from collapsing to mean predictions.

3. Key Contributions

1. Global 1m Resolution Map: The release of CHMv2, a global dataset (excluding Greenland/Antarctica) with 1m resolution, encoded in 213,109 Cloud-Optimized GeoTIFFs (COGs).
2. DINOv3 Integration: Demonstrating that DINOv3 SSL features significantly outperform DINOv2 and U-Net baselines for canopy height mapping, particularly in generalizing to new geographies.
3. Robust Registration Pipeline: A scalable, automated method for aligning optical and ALS data that mitigates the "chicken-and-egg" problem of needing aligned data to train a model for alignment.
4. Specialized Loss Formulation: A curriculum-based loss strategy that effectively balances structural fidelity (sharpness) with height accuracy, specifically addressing the bias in tall forests.

4. Results and Validation

The authors validated CHMv2 against independent ALS test sets, GEDI, and ICESat-2 observations.

• vs. CHMv1:
  • Accuracy: Mean Absolute Error (MAE) on the SatLidar v2 test set improved from 4.3m to 3.0m.
  • Correlation: $R^2$ improved from 0.53 to 0.86.
  • Bias: Significant reduction in underestimation of tall forests (≥30m).
  • Structure: Qualitative improvements in crown delineation, canopy gap characterization, and edge sharpness.
• vs. Other Global Products:
  • CHMv2 (MAE 3.0m) outperformed existing 10m and 30m global products (MAE 4.9m – 8.4m) when evaluated against the same ALS ground truth.
• vs. Spaceborne LiDAR (GEDI/ICESat-2):
  • GEDI: Global $R^2$ of 0.70, MAE 3.1m. Strong agreement across major biomes.
  • ICESat-2: Global $R^2$ of 0.60, MAE 2.9m.
  • The model shows consistent performance across tropical, temperate, and boreal regions, though accuracy decreases in areas with low sun elevation or heavy cloud/haze.

5. Significance and Impact

• Fine-Scale Monitoring: By providing 1m resolution, CHMv2 enables the derivation of structural metrics (gap fraction, edge density) previously only available in regions with ALS coverage. This is crucial for monitoring agroforestry, secondary forests, and degradation at sub-hectare scales.
• Climate and Carbon: The improved accuracy in tall forests and reduced bias directly enhances the reliability of above-ground biomass estimation and carbon accounting for climate finance (dMRV).
• Scalability: The methodology demonstrates that combining self-supervised vision transformers with rigorous data curation and automated registration can overcome the scarcity of high-quality labeled training data, offering a blueprint for future global geospatial products.

Limitations: The map is derived from single-date imagery (2017–2020), limiting temporal change detection. Accuracy can be affected by cloud/haze, low sun angles (shadows), and off-nadir viewing geometries. The model still slightly underestimates the upper tail of very tall emergent trees.