Using an artificial neural network, researchers at ETH Zurich have created the world’s first high-resolution vegetation height map for 2020 from satellite images. This map could provide key information for tackling climate change and species extinction, as well as for planning sustainable regional development.
Last year marked the start of the United Nations Decade of Ecosystem Restoration. This initiative aims to halt the degradation of ecosystems by 2030, to prevent it from progressing and, if possible, to repair the damage already caused. Carrying out this type of project requires precise foundations, such as surveys and maps of the existing vegetation.
In an interview, Ralph Dubayah, principal investigator of NASA’s Global Ecosystem Dynamics Investigation (GEDI) mission, explains: “We just don’t know how tall the trees are on a global scale. […] We need good world maps of tree locations. Because every time we cut down trees, we release carbon into the atmosphere, and we don’t know how much carbon we release.”
The precise analysis and preparation of this type of environmental data is the specialty of the EcoVision laboratory of the Department of Civil, Environmental and Geomatics Engineering at ETH Zurich. Founded in 2017 by Professor Konrad Schindler of ETH Zurich and Professor Jan Dirk Wegner of the University of Zurich, this lab is where researchers develop machine learning algorithms that enable automatic analysis of large-scale environmental data. scale. One of these researchers is Nico Lang. In his doctoral thesis, he developed a neural network-based approach to infer vegetation height from optical satellite images. Using this approach, he was able to create the first vegetation height map that covers the entire Earth: the Global Canopy Height Map.
The map’s high resolution is another first: Thanks to Lang’s work, users can zoom in up to 10 x 10 meters on any wood on Earth and check the height of trees. A forest survey of this type could show the way, especially when it comes to carbon emissions, because the height of trees is a key indicator of biomass and the amount of carbon stored. “About 95% of forest biomass is made up of wood, not leaves. Thus, biomass is strongly correlated with height,” explains Konrad Schindler, professor of photogrammetry and remote sensing.
Formed with laser scan data from space
But how does a computer read the height of a tree from a satellite image? “Since we don’t know what patterns the computer should look for to estimate height, we let it learn the best image filters on its own,” says Lang. It shows its neural network millions of examples, using images from the two Copernicus Sentinel-2 satellites operated by the European Space Agency (ESA). These satellites capture every location on Earth every five days with a resolution of 10×10 meters per pixel. These are the highest quality images currently available to the public.
The algorithm must also have access to the correct answer, i.e. the height of the tree derived from space laser measurements from NASA’s GEDI mission. “The GEDI mission provides scattered and globally distributed data on the height of vegetation between latitudes 51 degrees north and south, so the computer sees many different types of vegetation in the process of formation,” explains Lang. With the input and the response, the algorithm can itself acquire the filters for the textural and spectral patterns. Once the neural network is trained, it can automatically estimate vegetation height from the more than 250,000 images (about 160 terabytes of data) needed for the global map.
In specialist jargon, Lang’s neural network is known as a convolutional neural network (CNN). “Convolution” is a mathematical operation in which the algorithm drags a 3×3 pixel filter mask over the satellite image to obtain information about the brightness patterns in the image. “The trick here is that we stack the image filters. This gives the algorithm contextual information, since each pixel, from the previous convolution layer, already includes information about its neighbors,” Schindler explains. As a result, the EcoVision lab was the first to successfully use satellite maps to also reliably estimate tree heights up to 55 meters.
Because their many layers make these neural networks “deep”, this method is also called “deep learning”. About ten years ago he announced a major revolution in image processing. However, processing the amount of data remains very difficult: calculating the world map of vegetation height would take three years for a single powerful computer. “Fortunately, we have access to ETH Zurich’s high-performance computing cluster, so we didn’t have to wait three years for the map to be computed,” Lang says with a laugh.
Transparency by estimating uncertainties
Lang did not prepare a single CNN for this task, but several. This is called a set. “An important aspect for us was also to make users aware of the uncertainty of the estimate,” he says. The neural networks – five in total – were trained independently of each other, each returning its own estimate of tree height. “If all the models agree, then the answer is clear based on the training data. If the models arrive at different answers, it means there is greater uncertainty in the estimate,” explains Lang. Models also incorporate uncertainties into the data itself: if a satellite image is blurry, for example, the uncertainty is greater than when atmospheric conditions are good.
Foundation for Future Ecological Research
Thanks to its high resolution, Lang’s world map provides detailed information: “We have already discovered some interesting patterns,” says Schindler. “In the Rocky Mountains, for example, the forests are managed in fixed sections, and the rainforest also forms interesting structures that cannot be a coincidence.” Now, ecologists can interpret these globally captured patterns and data.
To allow this research to continue, the map and its source code will be made available to the public (see link). The first interested parties have already made contact: Walter Jetz, professor at Yale University, wants to use the Global Canopy Height Map for biodiversity modeling. However, the map could also be of interest to governments, administrations and NGOs. “Thanks to Sentinel-2, the height of vegetation can be recalculated every five days, which makes it possible to monitor the deforestation of the rainforest,” says Lang.
In addition, he adds, it is now also possible to globally validate regional findings, such as how canopies of tropical leaves act as a climate buffer. Coupled with the high carbon stock approach, which ranks forests according to their carbon storage and biodiversity value, the vegetation height map is an important basis for maintaining and enhancing ecosystems. According to Lang’s calculations, vegetation taller than 30 meters is found on only 5% of the landmass, and only 34% of it is in protected areas.
With the GEDI mission due to end in 2023, the new approach developed by Lang offers the possibility of continuing to map vegetation height in the future. However, extending the GEDI mission – something currently being discussed in the international media – is key to comparing its data with future satellite missions such as the ESA Biomass mission and calibrating the model for changes.
NASA Releases Groundbreaking Data on Forest Biomass-Carbon Products
Nico Lang, Walter Jetz, Konrad Schindler, Jan Dirk Wegner, A high resolution canopy height model of the Earth. arXiv:2204.08322v1 [cs.CV]arxiv.org/abs/2204.08322
Quote: Neural network can read tree height from satellite images (2022, April 20) Retrieved April 24, 2022 from https://techxplore.com/news/2022-04-neural-network-tree-heights -satellite.html
This document is subject to copyright. Except for fair use for purposes of private study or research, no part may be reproduced without written permission. The content is provided for information only.