Sub-Meter Canopy Height Synthesis from Sentinel-2 via Anchored Residual Flow
Organizations: Vibrant Planet · Arête Glaciers Initiative · Terradot
Abstract
Canopy height models (CHMs) - per-pixel maps of vegetation top height - are essential for forest carbon accounting, wildfire fuel mapping, and tree-level ecology, yet sub-meter CHMs require airborne LiDAR refreshed on multi-year national cycles. We produce 0.5 m CHMs from free 10 m Sentinel-2 imagery, a 20 cross-modal super-resolution task where optical reflectance must predict geometric height it does not directly observe. Motivated by the perception-distortion trade-off, we derive two operating points with shared conditioning. A deterministic regressor (Anchor) targets low distortion; a conditional flow matching model, trained on the Anchor's residual in a learned CHM latent space (AR-flow), targets perceptual fidelity. Trained on 197k LiDAR-paired tiles spanning 69 ecoregions across the contiguous United States and evaluated on 2,530 held-out tiles against eight baselines, the Anchor achieves 2.83 m MAE using only free satellite input. The AR-flow reduces sliced Wasserstein distance from the Anchor's 1.13 to 0.208, at a higher MAE of 3.68 m. Crown-scale statistics improve, but strict individual-tree detection remains weak: generated sub-pixel detail represents plausible structure rather than measured individual trees. Quarterly continental refresh becomes tractable, letting satellite monitoring inform timing and extent for more expensive data acquisitions.
Figures & tables
| Operating point | MAE (m) | RMSE (m) | EE | SWD | FWD ( ) | LPIPS | Input |
|---|---|---|---|---|---|---|---|
| Meta CHMv2 [ 5 ] | 2.90 | 4.35 | 0.380 | 1.31 | 11.06 | 0.395 | 0.6 m Maxar |
| Anchor (S2+DINOv3, ~197k tiles, 20 ep) | 2.83 | 4.06 | 0.326 | 1.13 | 3.47 | 0.436 | 10 m S2 |
| AR-flow (S2+DINOv3, ~197k tiles, 98 ep) | 3.68 | 5.43 | 0.370 | 0.208 | 2.32 | 0.391 | 10 m S2 |
| Product | Input (pixel size) | Native CHM grid | MAE (m) | RMSE (m) | EE | SWD | FWD ( ) | LPIPS |
|---|---|---|---|---|---|---|---|---|
| Meta CHMv2 [ 5 ] | Maxar Vivid2 (0.6 m) | 1 m | 2.90 | 4.35 | 0.380 | 1.31 | 11.06 | 0.395 |
| Ours Anchor | S2 (10 m) + DINOv3 | 0.5 m | 2.83 | 4.06 | 0.326 | 1.13 | 3.47 | 0.436 |
| Ours AR-flow | S2 (10 m) + DINOv3 | 0.5 m | 3.68 | 5.43 | 0.370 | 0.208 | 2.32 | 0.391 |
| GLAD [ 34 ] | Landsat (30 m) | 30 m | 4.08 | 5.48 | 0.613 | 2.13 | 19.17 | – |
| LandFire FCH [ 39 ] | Landsat composite (30 m) | 30 m | 4.70 | 6.30 | 0.563 | 2.58 | 21.83 | – |
| GEDI RH98 [ 13 ] | spaceborne LiDAR | sparse | 5.65 | 7.64 | 0.422 | 4.19 | 27.55 | – |
| Statistic | Anchor | AR-flow | Meta CHMv2 |
| Crown-count recovery (1.0 = LiDAR) | 0.21 | 0.78 | 0.82 |
| Crown-height KS vs full LiDAR field (lower better) | 0.171 | 0.059 | 0.152 |
| Tile-mean crown-area KS vs LiDAR (lower better) | – | 0.096 | 0.205 |
| Ripley departure, m (closer to 0) |
| Model | Recall | Precision | F1 |
|---|---|---|---|
| Anchor | 0.013 | 0.061 | 0.021 |
| AR-flow | 0.037 | 0.047 | 0.041 |
| Meta CHMv2 | 0.062 | 0.076 | 0.068 |
Appendix figures & tables33 assets
Supplementary material from the paper’s appendix.
Appendix
| Model | MAE (m) | EE | SWD | FWD ( ) |
|---|---|---|---|---|
| Stage 1 reference (full training) | 2.83 | 0.326 | 1.13 | 3.47 |
| AR-flow (residual target ) | 4.10 | 0.384 | 0.391 | 2.40 |
| Plain flow matching (direct target ) | 4.67 | 0.400 | 0.312 | 2.47 |
| MAE (m), lower better | A single | B four-season | (A B) | 95% CI |
|---|---|---|---|---|
| All pixels | 3.201 | 2.982 | 0.219 | |
| Vegetation ( m) | 5.591 | 5.284 | 0.307 |
| Operating point | MAE (m) | RMSE (m) | EE | SWD | FWD ( ) |
|---|---|---|---|---|---|
| AE roundtrip (LiDAR encode decode) | 0.79 | 1.33 | 0.148 | 0.46 | 1.04 |
| Anchor | 2.83 | 4.06 | 0.326 | 1.13 | 3.47 |
| AR-flow | 3.68 | 5.43 | 0.370 | 0.21 | 2.32 |
| Encoder | MAE (m) | EE |
|---|---|---|
| S2 + DINOv3 (ablation) | 3.259 | 0.476 |
| S2 + Clay v1.5 | 3.268 | 0.482 |
| S2 + AlphaEarth | 3.030 | 0.490 |
| S2 only (no foundation model) | 3.376 | 0.472 |
| DINOv3 only (no S2 latent) | 4.384 | 0.385 |
| Clay v1.5 only | 3.974 | 0.395 |
| Component | Architecture | Params | Input | Output | Frozen? |
|---|---|---|---|---|---|
| S2 encoder | DC-AE [ 7 ] , | ~12M | 12-band S2, | latent | yes |
| CHM autoencoder | DC-AE [ 7 ] , | ~50M | 1-ch CHM, | latent | yes |
| DINOv3 | ViT-L/16 SAT-493M [ 38 ] | ~300M | S2 RGB (B04/B03/B02), | 196 tokens, dim 1024 | yes |
| Stage 1 (Anchor) | UViT [ 2 ] | ~110M | S2 latent + DINOv3 x-attn | frozen for Stage 2 | |
| Stage 2 (AR-flow) | UViT [ 2 ] | ~110M | + DINOv3 x-attn | no |
| Paper | Cross-modal | Frozen MSE Stage-1 | Stage-2 transport |
|---|---|---|---|
| CorrDiff [ 27 ] | weak | yes | EDM-diffusion noise residual, -concat |
| ELIR [ 8 ] | no | co-trained L2 | consistency flow: noised L2 estimate clean (Tiny AE) |
| Kim and Chung [20] | yes (SAR optical) | no, source-cond | Brownian-bridge diffusion |
| C-DiffSET [ 11 ] | yes (SAR optical) | no | noise y diffusion |
| Ours | yes (optical height) | yes | AR-flow: linear-path noise residual, no -concat, CHM autoencoder latent |
| Product | Grid | tiles | Product MAE | Anchor MAE | AR-flow MAE |
|---|---|---|---|---|---|
| (control) Meta CHMv2 | 0.5 m | 2,530 | 2.892 | 2.825 | 3.673 |
| Meta CHMv2 | 1 m | 2,530 | 2.816 | 2.712 | 3.580 |
| ETH Global | 10 m | 2,527 | 6.845 | 1.838 | 2.440 |
| Pauls / AI4Forest | 10 m | 2,530 | 6.079 | 1.839 | 2.444 |
| EchoSat | 10 m | 2,530 | 5.880 | 1.839 | 2.444 |
| GLAD | 30 m | 2,526 | 3.102 | 1.367 | 1.714 |
| Quarter | MAE (m) | RMSE (m) | EE | MAE vs JJA | |
|---|---|---|---|---|---|
| JJA (dominant acquisition season) | 789 | 2.75 | 4.13 | 0.321 | – |
| MAM (spring) | 789 | 4.07 | 5.61 | 0.323 | |
| SON (autumn) | 789 | 3.99 | 5.39 | 0.324 | |
| DJF (winter) | 789 | 4.64 | 6.28 | 0.326 |
| Anchor | AR-flow | Meta CHMv2 | |||||||
|---|---|---|---|---|---|---|---|---|---|
| IoU | R | P | F1 | R | P | F1 | R | P | F1 |
| 0.2 | 0.178 | 0.689 | 0.283 | 0.426 | 0.480 | 0.452 | 0.512 | 0.566 | 0.537 |
| 0.3 | 0.085 | 0.401 | 0.141 | 0.236 | 0.296 | 0.262 | 0.321 | 0.385 | 0.350 |
| 0.4 | 0.038 | 0.180 | 0.063 | 0.109 | 0.137 | 0.121 | 0.163 | 0.196 | 0.178 |
| 0.5 | 0.013 | 0.061 | 0.021 | 0.037 | 0.047 | 0.041 | 0.062 | 0.076 | 0.068 |
| Measure | Question it answers | What it cannot see |
|---|---|---|
| RAPSD | Is variance present at crown scale? | Placement. A field with correct texture in the wrong locations scores perfectly. |
| Crown-height / crown-size KS | Are the marginal distributions right? | Placement, and any joint structure between size and location. |
| Ripley | Is the spatial arrangement right? | Height. Two fields with identical point patterns may have unrelated canopy heights. |
| SWD / FWD / LPIPS | Does the field match reference patch statistics or features? | Geometric correctness; perceptual agreement does not guarantee accurate crown placement. |
| ITCD recall / precision | Can discrete objects be recovered? | Everything below the matching threshold, and the threshold is a free parameter. |
| Block R 2 | Is fine-scale pixel variation right? | Perceptual plausibility; it penalises correct-looking detail placed imperfectly. |
| Method | $/km 2 | CONUS | Global | Cadence | Access |
|---|---|---|---|---|---|
| Airborne LiDAR (3DEP) | $100–300 | $0.8–2.4 B | $15–45 B | multi-year | open |
| Maxar Vivid2 + CHMv2 | commercial | commercial | commercial | annual | gated by license |
| Ours (forest-extent) | $0.01 | $6–9K | $75–115K | quarterly | open |
| tile_id | HiForm acres | damage (m) | out (m) | damage out (m) |
|---|---|---|---|---|
| 6440_19521 | 59.7 | |||
| 6587_19647 | 54.6 | |||
| 6535_19561 | 35.6 | |||
| mean | 50.0 |
| Model | Reference truth | tiles | MAE (m) | RMSE (m) | Pearson | MAE vs 3DEP (m) |
|---|---|---|---|---|---|---|
| Anchor | NEON airborne (~1 m) | 8 | ||||
| AR-flow | NEON airborne (~1 m) | 8 | ||||
| Anchor | GEDI RH98 spaceborne (25 m) | 2,526 | ||||
| AR-flow | GEDI RH98 spaceborne (25 m) | 2,526 |
| Scale | Anchor vs GEDI | AR-flow vs GEDI | GEDI vs 3DEP |
|---|---|---|---|
| Per-pixel (pooled) | 0.43 | 0.38 | 0.40 |
| 25 m footprint disc | 0.44 | 0.43 | 0.46 |
| Coarse 100 m | 0.65 | 0.65 | 0.68 |
| Coarse 500 m | 0.87 | 0.87 | 0.89 |
| Meta-MAE group | n ecos | n tiles | Anchor winrate (MAE) | Mean MAE (m) |
|---|---|---|---|---|
| Low error (Meta MAE m) | 14 | 766 | 3/14 | (Meta wins) |
| Medium error ( to m) | 22 | 682 | 6/22 | (Meta wins) |
| High error (Meta MAE m) | 33 | 1,094 | 18/33 | (Anchor wins) |
| Ecoregion | L2 biome | LiDAR (m) | Anchor MAE (m) | Class | |
|---|---|---|---|---|---|
| Coast Range | Marine West Coast Forest | 30 | 13.9 | 6.12 | meaningful |
| Klamath Mtns / CA High N. Coast | Western Cordillera | 48 | 15.4 | 6.30 | meaningful |
| Northern Rockies | Western Cordillera | 40 | 10.2 | 5.26 | meaningful |
| Interior Plateau | SE USA Plains | 39 | 9.8 | 5.18 | meaningful |
| Idaho Batholith | Western Cordillera | 46 | 6.2 | 4.55 | meaningful |
| Middle Rockies | Western Cordillera | 39 | 5.4 | 4.18 | meaningful |
| Model | Block R 2 | valid |
|---|---|---|
| Anchor | ||
| AR-flow | ||
| Meta CHMv2 |
| Product | Input imagery | Nominal vintage |
|---|---|---|
| Our pipeline (Anchor / AR-flow) | Sentinel-2 L2A | year-aligned (every tile) |
| Meta CHMv2 [ 5 ] | Maxar Vivid2 | 2017–2020 (~80% 2018–2020) |
| ETH-S2 [ 21 ] | Sentinel-2 (multi-date, fused) | ~2020 |
| GLAD [ 34 ] | Landsat | ~2019 |
| LandFire FCH [ 39 ] | Landsat, lidar and plot-trained models | LF 2023 update (disturbance through 2023) |
| GEDI RH98 [ 13 ] | GEDI (~25 m footprints) | 2019 onward (sparse) |
| Full cohort | LiDAR 2019–2020 | ||||
|---|---|---|---|---|---|
| Product | MAE (m) | MAE (m) | RMSE (m) | ||
| Anchor | 2,530 | 2.820 | 911 | 2.680 | 3.800 |
| AR-flow | 2,530 | 3.664 | 911 | 3.486 | 5.105 |
| Meta CHMv2 | 2,530 | 2.896 | 911 | 2.898 | 4.259 |
| GLAD | 2,527 | 4.083 | 908 | 3.974 | 5.184 |
| LandFire FCH | 2,530 | 4.709 | 911 | 4.832 | 6.277 |
| MAE (m) | RMSE (m) | EE | |
|---|---|---|---|
| 0.0 (pure Anchor) | 2.83 | 4.06 | 0.557 |
| 0.2 | 2.86 | 4.11 | 0.458 |
| 0.4 | 2.98 | 4.29 | 0.381 |
| 0.6 | 3.15 | 4.59 | 0.334 |
| 0.8 | 3.39 | 4.98 | 0.306 |
| 1.0 (pure AR-flow) | 3.68 | 5.43 | 0.289 |
| Statistic | Value |
|---|---|
| Mean per-pixel std | 2.28 m |
| Pearson ( , error ) | 0.391 |
| Spearman ( , error ) | 0.414 |
| Empirical coverage ( error ) | 65.7% |
| Tile | Region / forest type | GT mean (m) | AR-flow MAE (m) |
| Saltdal | Norway, boreal pine/birch | 4.0 | 4.29 |
| Steinkjer | Norway, boreal spruce | 3.9 | 4.03 |
| Ringerike | Norway, mixed boreal | 4.1 | 7.74 |
| Elverum | Norway, boreal spruce/pine | 4.9 | 5.67 |
| 092g003 | BC, coastal Douglas-fir | 4.6 | 2.12 |
| 082f008 | BC, interior cedar-hemlock | 7.3 | 6.80 |
| code | ecoregion | code | ecoregion | ||
|---|---|---|---|---|---|
| 1 | Coast Range | 79 | 44 | Nebraska Sand Hills | 0 |
| 2 | Puget Lowland | 31 | 45 | Piedmont | 24 |
| 3 | Willamette Valley | 0 | 46 | Northern Glaciated Plains | 0 |
| 4 | Cascades | 35 | 47 | Western Corn Belt Plains | 0 |
| 5 | Sierra Nevada | 96 | 48 | Lake Agassiz Plain | 6 |
| 6 | Central California Foothills and Coastal Mountains | 77 | 49 | Northern Minnesota Wetlands | 6 |