Skip to main content
Back to timeline
arXivSource publication:

Free Sentinel-2 Imagery Yields 0.5 m Canopy Heights: Anchor Reaches 2.83 m Mean Error on 2,530 Held-Out Tiles

Synopsis

Canopy height maps sampled at 0.5 m can now be produced from free 10 m Sentinel-2 imagery, where sub-meter products previously depended on airborne LiDAR refreshed on multi-year national cycles. Two outputs share the same conditioning: the deterministic regressor Anchor reaches 2.83 m mean absolute error on 2,530 held-out tiles, slightly below Meta CHMv2's 2.90 m from commercial Maxar imagery, while the generative AR-flow, trained on the Anchor's residual, cuts sliced Wasserstein distance from 1.13 to 0.208 at a higher 3.68 m mean absolute error. The authors stress that 0.5 m refers to output sampling only; individual-tree detection stays weak, and generated sub-pixel detail is plausible structure rather than measured trees.

Source-provided article image: Sub-Meter Canopy Height Synthesis from Sentinel-2 via Anchored Residual Flow
arXiv

Interpretation

Free 10 m Sentinel-2 imagery can now produce canopy height maps sampled at 0.5 m, with two outputs aimed at different uses: one for height accuracy, one for canopy structure appearance. Sub-meter canopy height maps previously depended on airborne LiDAR recollected on multi-year national cycles; the strongest commercial-input baseline in the evaluation, Meta CHMv2, uses 0.6 m Maxar imagery. On 2,530 year-aligned held-out tiles, the Anchor reaches 2.83 m mean absolute error against Meta CHMv2's 2.90 m, with a paired-bootstrap 95% confidence interval excluding zero; AR-flow lowers sliced Wasserstein distance from the Anchor's 1.13 to 0.208 at 3.68 m mean absolute error.

Both outputs share one conditioning bundle; they differ only in whether a second stage samples a generative residual on top of the Anchor. A deterministic regressor returns only the conditional mean and smooths crown-scale detail, while a generative model supplies the latent structural variation that regression discards. The model trains on roughly 197k LiDAR-paired tiles spanning 69 ecoregions of the contiguous United States, conditioned on 12-band Sentinel-2 composites and frozen DINOv3 features, and operates in the latent space of a frozen canopy height autoencoder.

0.5 m sampling is not 0.5 m structure, and the two outputs differ at crown scale far more than mean error suggests. In a 16-tile spectral probe, the Anchor retains 0.008 of LiDAR's structural power at the 10 m input scale, against 0.195 for Meta CHMv2 and 0.95 for AR-flow. Individual-crown detection at IoU 0.5 gives F1 of 0.021 for the Anchor, 0.041 for AR-flow and 0.068 for Meta CHMv2, all near the floor; AR-flow under-clusters relative to LiDAR, so its detail is synthesized rather than recovered.

Perspective

The demonstrated scope is year-aligned held-out tiles across 69 ecoregions of the contiguous United States, with 0.5 m output sampling; the authors state that 0.5 m refers to output sampling, not error magnitude, and that sub-10 m structure is context-conditioned synthesis rather than measurement. Model weights remain withheld as institutional IP, and the planned release covers predictions, LiDAR references and per-tile metrics.

The authors report that AR-flow's injected high-frequency detail lowers Block R2 and that it under-clusters relative to LiDAR, so clumping estimates derived from it lean toward regularity; winter leaf-off raises error about 1.7-fold, which is why the operational cadence is spring, summer and autumn. NEON cross-validation covers only 8 tiles, tropical and mangrove biomes remain unevaluated, and uncertainty intervals are not yet calibrated.