Scientists at ETH Zurich’s EcoVision Lab have developed a deep learning framework to map treetop height globally at high resolution, using publicly available optical satellite images as input. The Global Canopy Height Map is the first map of its kind. It could become a critical tool for tracking carbon emissions that contribute to climate change and for planning sustainable regional development. To acquire the data needed for the first global canopy height map, EcoVision Lab researchers drew from two sources: NASA’s Global Ecosystem Dynamics Investigation (GEDI) and the Copernicus Sentinel-2 satellites operated by the European Space Agency. GEDI, which has the highest resolution and densest sampling of any lidar ever put in orbit, makes laser-ranging observations of nearly all tropical and temperate forests on Earth and provides high-resolution measurements of Earth’s 3D structure. It delivers sparse but well-distributed data about canopy height worldwide. Optical satellite images from Sentinel-2 satellites capture every location on Earth every five days with a resolution of 10 × 10 meter/pixel. Sentinel-2 satellites offer dense observations globally, though they cannot enable measurement of vertical structures. By combining data from GEDI with data from Sentinel-2, the researchers created a probabilistic deep learning model to retrieve canopy height from Sentinel-2 images anywhere on Earth, and to quantify the uncertainty in these estimates with data from GEDI. The researchers prepared an ensemble of convolutional neural networks to map tree height globally. The neural networks were shown millions of sample images from two Sentinel-2 satellites. “Since we don’t know which patterns the computer needs to look out for to estimate height, we let it learn the best image filters itself,” said researcher Nico Lang, who developed the networks. The neural networks’ algorithm accesses the appropriate answer to tree height from the space laser measurements taken by GEDI. “The GEDI mission delivers globally distributed, sparse data on the vegetation height between the latitudes of 51° north and south, so the computer sees many different vegetation types in the training process,” Lang said. Researchers at ETH Zurich have developed a world map that uses machine learning to derive vegetation heights from satellite images in high resolution. Courtesy of EcoVision Lab. With input from the optical satellite images and the lidar measurements, the algorithm is able to acquire the filters for textural and spectral patterns on its own. By sliding a 3- × 3-pixel filter mask over the satellite image, the algorithm obtains information on brightness patterns in the image. “The trick here is that we stack the image filters,” professor Konrad Schindler said. “This gives the algorithm contextual information, since every pixel, from the previous convolution layer, already includes information about its neighbors.” Five neural networks were trained independent of each other, with each returning its own estimate of tree height. “If all the models agree, then the answer is clear based on the training data,” Lang said. “If the models arrive at different answers, it means there is a higher uncertainty in the estimate.” The models also incorporate uncertainties that can arise from the input data; for example, when a satellite image is hazy, the uncertainty is greater than when atmospheric conditions are clear. Once trained, the neural networks require only image data, meaning that the map can be updated annually with satellite images from Sentinal-2. At the same time, the longer the GEDI mission collects data, the denser the reference data for the Global Canopy Height Map will be. Once a neural network has been trained, it can automatically estimate the vegetation height from the more than 250,000 images — about 160 TB of data — needed for the global map. According to the researchers, calculating the global vegetation height map would take a single powerful computer three years. “Fortunately, we have access to the ETH Zurich high-performance computing cluster, so we didn’t have to wait three years for the map to be calculated,” Lang said. The Global Canopy Height Map can provide insight into carbon emissions, as tree height is a key indicator of biomass and the amount of carbon stored. “Around 95% of the biomass in forests is made up of wood, not leaves. Thus, biomass strongly correlates with height,” Schindler said. According to the map, only 5% of the global landmass is covered by trees taller than 30 m, and only 34% of these tall canopies are located within protected areas. The ETH Zurich model will enable consistent, uncertainty-informed worldwide mapping and will support ongoing monitoring to detect change and inform decision-making. It can serve ongoing efforts in forest conservation, and it has the potential to foster advances in climate, carbon, and biodiversity modeling. The global map and underlying source code and models are being made publicly available to support conservation efforts. For more information, visit N. Lang et al., “A high-resolution canopy height model of the Earth” (www.doi.org/10.48550/arXiv.2204.08322).
