High Resolution Canopy Height Maps

A canopy height map gives, for each patch of ground, how tall the trees on it stand. Measuring it directly takes lidar, a laser scanner flown on a plane, and good aerial lidar does not exist for most of the world.³ Meta's researchers, working with the World Resources Institute (WRI), trained a model to estimate tree height from ordinary colour satellite photographs at half a metre per pixel, sharp enough to pick out single trees.^2,3

The model learned in two stages. It first studied about 18 million satellite images with no labels at all, using Meta's DINOv2 method on pictures chosen to show mostly vegetation.² It was then taught height from aerial lidar covering a few thousand square kilometres of US forest.² The satellite images were Maxar's, from 2017 to 2020.⁴ The paper, published in Remote Sensing of Environment in 2024, reports an average error of 2.8 metres.⁴ The model also works on photographs from planes or drones at the same resolution.³

The first maps covered California and the state of São Paulo in Brazil, released in April 2023.³ A global map and one for sub-Saharan Africa have followed.¹ All of them sit in a public Amazon S3 bucket as GeoTIFF files, each with a GeoJSON file giving the date of the imagery underneath, and Meta and WRI run a viewer for them on Google Earth Engine.^1,2

Tree height is a step towards how much carbon a forest holds, and a map at the scale of single trees can follow small woodlots and trees on farms as well as whole forests.³ A model of a watershed, a city's heat or a fire landscape could take it as its canopy layer. Each height is an estimate from a photograph taken on a particular day, and the date files say which.¹

The maps are under CC BY 4.0, free for any use with credit to Meta and WRI and to Maxar for the imagery.¹ The inference code and the trained weights are under Apache 2.0 on GitHub; the correction against GEDI, a space-borne lidar, that the paper applied to its maps is not part of that release.^2,4