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

TL;DR

VibrantSR targets the problem of obtaining high resolution canopy height models with regular update capability by deriving 0.5 m CHMs from freely available 10 m Sentinel-2 imagery. It learns a latent flow mapping that translates compressed Sentinel-2 representations into CHM latents, modeled as a transport between fixed latent manifolds using a velocity field in an ODE framework. Across 22 EPA Level-3 eco-regions in the western United States, it achieves a MAE of 4.39 m for heights >= 2 m, outperforming several satellite-based baselines and preserving fine-scale canopy variability. While not matching aerial imagery based approaches in absolute accuracy, VibrantSR enables scalable, seasonal monitoring and carbon accounting at continental scales, with plans to incorporate multi-date inputs and additional modalities to further boost performance.

Abstract

We present VibrantSR (Vibrant Super-Resolution), a generative super-resolution framework for estimating 0.5 meter canopy height models (CHMs) from 10 meter Sentinel-2 imagery. Unlike approaches based on aerial imagery that are constrained by infrequent and irregular acquisition schedules, VibrantSR leverages globally available Sentinel-2 seasonal composites, enabling consistent monitoring at a seasonal-to-annual cadence. Evaluated across 22 EPA Level 3 eco-regions in the western United States using spatially disjoint validation splits, VibrantSR achieves a Mean Absolute Error of 4.39 meters for canopy heights >= 2 m, outperforming Meta (4.83 m), LANDFIRE (5.96 m), and ETH (7.05 m) satellite-based benchmarks. While aerial-based VibrantVS (2.71 m MAE) retains an accuracy advantage, VibrantSR enables operational forest monitoring and carbon accounting at continental scales without reliance on costly and temporally infrequent aerial acquisitions.

Figures (8)

• Figure 1: Spatial partitioning of training and validation tiles across the western United States. The left panel shows the distribution of tiles assigned to the training (blue) and validation (red) folds. The right panel provides a regional zoom illustrating strict spatial separation between folds.
• Figure 2: Error metrics across EPA Level 3 ecoregions.
• Figure 3: VibrantSR architecture showing the flow from Sentinel-2 inputs through frozen autoencoders and the trainable flow matching network to CHM outputs.
• Figure 4: Mean absolute error (MAE) as a function of output resolution for VibrantSR and baseline CHM products. Box–whisker plots summarize tile-level error at 0.5 m, 10 m, and 30 m. VibrantSSR maintains stable accuracy across resolutions.
• Figure 5: (Top) Mean absolute error (MAE) by EPA Level-3 ecoregion for VibrantSR and competing CHM products. Each box–whisker summarizes the distribution of tile-level MAE within an ecoregion at 0.5 m resolution. VibrantSR is consistently among the lowest-error methods in structurally complex, high-biomass regions (e.g., Rockies, Sierra Nevada, Cascades), while performance converges in flatter and sparsely vegetated regions. (Bottom) Height-binned distribution of per-pixel canopy-height error ($\hat{h}-h$) at 0.5 m resolution. Boxes show the interquartile range, horizontal lines indicate the median, and whiskers denote the 5th--95th percentiles; grey bars give the relative lidar sample count. VibrantSR shows a smooth shift toward increasing underestimation with height, whereas LANDFIRE and ETH exhibit larger and less stable biases, particularly at low and very high canopy heights.