Researchers at ETH Zurich have developed the first high-resolution worldwide vegetation height map for 2020 using satellite pictures and an artificial neural network. This map could be useful in the fight against climate change and species extinction, as well as in planning for long-term regional development.
The United Nations Decade on Ecosystem Restoration began last year. This program aims to halt environmental degradation by 2030, prevent it from continuing, and, if feasible, repair the damage that has already been done. These kinds of projects necessitate precise foundations, such as surveys and maps of current vegetation.
“We simply do not know how tall trees are globally. We need good global maps of where trees are. Because whenever we cut down trees, we release carbon into the atmosphere, and we don’t know how much carbon we are releasing.”
Ralph Dubayah, the Principal Investigator of NASA’s Global Ecosystem Dynamics Investigation (GEDI)
In an interview, Ralph Dubayah, NASA’s Principal Investigator for the Worldwide Ecosystem Dynamics Investigation (GEDI) mission, explains in an interview: “We simply have no idea how big trees are on a global scale. […] We require accurate global maps of tree locations. Because we release carbon into the atmosphere when we take down trees, and we don’t know how much carbon we’re releasing.
The EcoVision Lab of the ETH Zurich Department of Civil, Environmental, and Geomatic Engineering specializes in analyzing and preparing precisely this type of environmental data. This lab was founded in 2017 by ETH Zurich Professor Konrad Schindler and University of Zurich Professor Jan Dirk Wegner to create machine learning algorithms that enable automatic analysis of large-scale environmental data. Nico Lang is one of these scientists. In his doctoral thesis, he created a neural network-based approach for determining vegetation height from optical satellite pictures. He was able to generate the first vegetation height map that covered the entire Earth using this method: the Global Canopy Height Map.
Another first is the map’s high resolution. Thanks to Lang’s work, viewers may zoom in to as little as 10×10 meters on every piece of woodland on Earth and verify tree height. A forest census like this could pave the way forward, especially in terms of reducing carbon emissions, because tree height is a major indicator of biomass and carbon storage. “Wood, not leaves, makes up nearly all of the biomass in forests. As a result, biomass has a strong relationship with height, “Professor of Photogrammetry and Remote Sensing Konrad Schindler explains.
Laser scanning data from space has been used to train
But how does a computer determine the height of trees from a satellite image? Because we don’t know which patterns the computer should be looking for to estimate height, “we let it discover the optimum picture filters on its own.” He uses photos from the European Space Agency’s two Copernicus Sentinel-2 satellites to show his neural network millions of examples (ESA). Every five days, these satellites photograph every location on Earth at a resolution of 10×10 meters per pixel. They are the highest-resolution photos currently available to the general public.
The algorithm must also have access to the right answer, which is the tree height derived from NASA’s GEDI mission’s space laser measurements. “The GEDI mission provides globally spread, sparse data on vegetation height between 51 degrees north and south latitudes,” Lang continues, “so the computer sees many distinct vegetation kinds in the training process.” The algorithm may create its own filters for textural and spectral patterns using the input and answer. After training, the neural network can automatically estimate vegetation height from the more than 250,000 photos (about 160 gigabytes of data) required for the worldwide map.
Lang’s neural network is referred to as a “convolutional neural network” in technical terms (CNN). “Convolution” is a mathematical operation in which the algorithm slides a 3×3 pixel filter mask over the satellite picture in order to gain information about the image’s brightness patterns. The trick is that the picture filters are stacked in this case. Because every pixel from the preceding convolution layer already contains information about its neighbors, this provides context to the algorithm “according to Schindler.” As a result, the EcoVision Lab was the first to correctly predict tree heights of up to 55 meters using satellite data.
This method is also known as “deep learning” because of the multiple layers that make these neural networks “deep.” Ten years ago, it signaled a fundamental shift in image processing. However, dealing with the sheer volume of data remains difficult: generating the worldwide vegetation height map would take three years on a single powerful computer. Fortunately, we have access to the ETH Zurich high-performance computing cluster, so we didn’t have to wait three years to generate the map, “Lang jokes.”
Transparency is achieved by estimating uncertainty.
Lang did not prepare a single CNN for this assignment, but rather numerous. This is referred to as an ensemble. “A key part for us was also letting users know the estimate’s uncertainty,” he explains. The five neural networks were trained independently of one another, with each returning its own estimate of tree height. Based on the training data, if all of the models agree, the answer is obvious. If the models produce different results, it indicates that the estimate is more uncertain, “Lang elaborates. The models also account for uncertainties in the data: if a satellite image is cloudy, for example, the uncertainty is higher than when air conditions are favorable.
The groundwork for future ecological studies
Lang’s worldwide map provides extensive information due to its high resolution: “We have already uncovered interesting patterns,” Schindler explains. “Forests, for example, are controlled in fixed parts of the Rocky Mountains, and the rainforest produces remarkable shapes that cannot be coincidence.” Ecologists may now analyse these gathered patterns and data on a global scale.
The map and its source code will be made freely available to allow this research to continue (see link). The first interested parties have already contacted us: Yale University professor Walter Jetz intends to use the Global Canopy Height Map to model biodiversity. The map, however, may be of interest to governments, administrative entities, and non-governmental organizations (NGOs). “With Sentinel-2, vegetation height can be recalculated every five days, allowing us to track rainforest degradation,” Lang explains.
Furthermore, he says, localized findings may now be globally validated, such as how tropical leaf canopies operate as a climatic buffer. The vegetation height map, when combined with the High Carbon Stock Approach, which classifies forests based on their carbon storage and biodiversity value, is an important basis for preserving and enhancing ecosystems. According to Lang’s calculations, vegetation with a height of more than 30 meters can be found on only 5% of the landmass, with only 34% of it placed in protected zones.
With the GEDI mission expected to finish in 2023, Lang’s newly established technique allows for future mapping of vegetation height. However, extending the GEDI mission—which is now being discussed in the worldwide media—is critical for comparing its data with future satellite missions such as the ESA Biomass project and calibrating the model for changes.
