Sub-Meter Canopy Height from Sentinel-2 via Anchored Residual Flow

Evaluation data and model outputs for "Sub-Meter Canopy Height from Sentinel-2 via AnchoredResidual Flow" (NeurIPS 2026, Main Track). Everything needed to check the paper's headlineaccuracy against public data. ## Where the data lives This record holds the manifests, score tables and documentation. The rasters exceed Zenodo'sper-record limit and are hosted on public object storage instead, world-readable over plain HTTPSwith no credentials and no requester-pays: https://vp-open-science.s3.us-west-2.amazonaws.com/projects/vibrant-sr/2026/neurips-eval-bench/conus/v1.0.0/ You do not need to build those paths by hand. Every row of the manifest carries a direct link toeach of its three rasters, as anchor_url / ar_flow_url / groundtruth_url, plus matching _s3columns for tooling that prefers object-store URIs. ## Files in this record - manifest.csv and manifest.parquet - one row per tile: geometry, split, EPA level-3 ecoregion, Sentinel-2 and LiDAR acquisition years, a scoreable flag, and raster URLs- per_tile_scores.csv - MAE per tile for both arms, with valid-pixel counts- per_ecoregion_scores.csv - MAE by ecoregion- README.md - raster conventions and reproduction steps- LICENSE, CITATION.cff ## The cohort 2,542 tiles of 500 x 500 m at 0.5 m ground sample distance, EPSG:6931, spanning 69 EPA level-3ecoregions across the contiguous United States. Each tile carries two model predictions, aStage-1 MMSE anchor and a Stage-2 anchored residual flow, alongside airborne LiDAR canopy heightderived from USGS 3DEP. The Sentinel-2 composite and the LiDAR acquisition are year-aligned pertile. 2,530 tiles are scoreable. Twelve have no valid ground-truth pixel and carry scoreable = false;they ship anyway for completeness. Filter on that column rather than hardcoding a count. Predictions are float32 metres with NaN nodata. Ground truth is int16 centimetres with nodata32767. Predictions carry a 20-pixel nodata margin, so the scored region is the central 960 x 960window. Score only where the ground truth is valid AND every arm is finite, so all arms arecompared on identical pixels. ## Reproducing the headline Cohort MAE is a tile-weighted mean across the 69 ecoregions, not a pooled per-pixel mean over thewhole set. The two differ, and the paper reports the former. Compute per-tile MAE, average withineach ecoregion, then weight those means by tile count. - Anchor: 2.8250 m (paper reports 2.83)- Anchored residual flow: 3.6734 m (paper reports 3.677) Verified over 2,530 tiles and 2.26e9 pixels. per_tile_scores.csv and per_ecoregion_scores.csvhold the intermediate values, so each step can be checked independently rather than taken ontrust. Mean ground-truth canopy height across the cohort is 5.82 m, which is the MAE a zero predictorwould achieve. Both arms sit well below it. ## Reading the numbers The residual-flow arm has HIGHER MAE than the anchor by design. It trades pixel accuracy forperceptual realism, which is the distortion-perception trade-off the paper is built around. Usethe anchor where pixel accuracy matters, such as carbon and biomass work that integrates overarea, and the flow arm where structural realism matters, such as crown delineation and fuelmapping. MAE varies by an order of magnitude across ecoregions, from roughly 11 m in tall closed-canopyforest to under 0.2 m in near-treeless country. A cohort mean of 2.83 m is therefore shaped asmuch by cohort composition as by model skill, and should not be read as the error a user shouldexpect in forest. Use per_ecoregion_scores.csv for the region that matters to you. ## Companion raster release Predicted canopy height over 12,600 km2 of Western conifer forest is published separately and isalready available: https://vp-open-science.s3.us-west-2.amazonaws.com/projects/vibrant-sr/2026/neurips-raster-release/west/v1.0.0/ 56 tiles of 15 x 15 km at 0.5 m, anchored residual flow only, from June-August 2025 Sentinel-2composites, selected where every sampled airborne-LiDAR reference predates 2019. A static STACcatalogue and a per-tile parquet manifest accompany the rasters. There is no ground truth on thatfootprint, so those tiles cannot be scored; this record is where the accuracy figures come from. ## Not included Model weights are not released. Training and evaluation code are not included; both model inputsare public archives (Copernicus Sentinel-2, USGS 3DEP), so the corpus is rebuildable from opensources even though the tooling is not published. Third-party baseline rasters (Meta CHMv2, ETH,GLAD, LandFire, EchoSat, Pauls/AI4Forest, Simard, GEDI) are cited in the paper and remainavailable from their original sources; they are not redistributed here. ## Licence CC-BY-4.0. Ground-truth derivatives are from USGS 3DEP, a United States Government work in thepublic domain.