CHMv2 Full-Image Tiled Inference

Canopy height estimation over large satellite images using CHMv2 (Meta, March 2026).

Model

facebook/dinov3-vitl16-chmv2-dpt-head — DINOv3 ViT-L backbone + DPT depth head, 0.3B parameters.

Input

• 20000×20000px Google satellite image, 0.25m/px resolution, EPSG:32555
• 4-band RGBA uint8 GeoTIFF

Scripts

test_chmv2.py

Runs inference on a single 512×512 crop from the center of the image. Good for quickly testing the environment.

run_chmv2_full.py

Processes the full 20000×20000 image using overlapping tiles with Hann-window blending:

• Tile size: 512×512 (max that fits in GPU fp32 with ~5.6GB VRAM)
• Overlap: 64px each side, stride 384px → 2704 tiles
• Blending: 2D Hann weight mask to eliminate seam artifacts
• Memory: np.memmap accumulators to avoid RAM OOM
• Output: Compressed tiled GeoTIFF with original CRS and geotransform

Runtime: ~13 minutes on a GPU with 5.6GB VRAM.

Results

Stat	Value
Min height	0.01 m
Max height	68.46 m
Mean height	11.45 m

Setup

conda activate tree_height_env
pip install rasterio tqdm
pip install git+https://github.com/huggingface/transformers  # dev ≥5.3.0 required

How tiling works

See TILING_EXPLAINED.md for a detailed explanation of the overlap and blending strategy.