VibrantSR: Sub-Meter Canopy Height Models from Sentinel-2 Using Generative Flow Matching

Kiarie Ndegwa

TL;DR: We use generative flow matching to super-resolve 10 m Sentinel-2 imagery into 0.5 m canopy height maps, outperforming existing satellite-derived products.

Abstract: We present VibrantSR, a generative super-resolution framework that estimates 0.5 m canopy height models (CHMs) from 10 m Sentinel-2 imagery using latent-space flow matching. Evaluated across 22 EPA Level 3 eco-regions in the western United States, VibrantSR achieves a Mean Absolute Error of 4.39 m for canopy heights ≥2 m, outperforming satellite-based benchmarks Meta (4.83 m), LANDFIRE (5.96 m), and ETH (7.05 m) with lower edge error. An aerial-imagery baseline (2.71 m MAE) retains an accuracy advantage, reflecting the fundamental resolution gap between 10 m and 0.5 m inputs. VibrantSR offers a scalable alternative for regional forest monitoring that improves on existing satellite-derived CHM products while avoiding reliance on costly aerial acquisitions.

Submission Number: 21