AI for Conservation: Training Artificial Intelligence to Identify Trees from Satellite Imagery

Subscribe To Download This Insight

2Q 2021 | IN-6159

Training AI to identify trees from satellite imagery will help climate change prevention efforts.

Registered users can unlock up to five pieces of premium content each month.

Log in or register to unlock this Insight.

 

Learning to Spot Trees

NEWS


Scientists from NASA’s Goddard Space Flight Center in Greenbelt, Maryland, along with international collaborators, have demonstrated a new method for mapping the location and size of trees growing outside of forests. Billions of trees were discovered in arid and semi-arid regions, laying the groundwork for more accurate global measurement of carbon storage on land. Using supercomputers and machine learning algorithms, the team mapped the crown diameter — the width of a tree when viewed from above — of more than 1.8 billion trees across an area of more than 500,000 square miles, or 1,300,000 square kilometers. The team mapped how tree crown diameter, coverage, and density varied depending on rainfall and land use.

The scientists used high-resolution commercial satellite imagery from DigitalGlobe to identify and measure the individual trees. These images came from the commercial GeoEye-1, QuickBird-2, WorldView-2, and WorldView-3 satellites with the team focusing the study area on the dryland regions of West Africa, consisting of varying landscapes — arid, semi-arid, and humid sub-tropics. The team ran a powerful computing algorithm called a fully convolutional neural network, also known as “deep learning”, on the University of Illinois’ Blue Waters, one of the world’s fastest supercomputers. This involved the team having to first spend a year manually marking almost 90,000 individual trees across a range of terrain, creating a training dataset for the model and allowing it to “learn” which shapes and shadows indicated the presence of trees.

Shining a Light on Outside Forests

IMPACT


One of the key causes of climate change is the rising CO2 emissions from human activities, including the burning of trees and fossil fuels among many others. According to the International Energy Agency (IEA), the global energy-related CO2 emissions fell by 5.8% in 2020, largely due to the COVID-19 pandemic. However, despite this drop in emissions, countries around the globe still have some ways to go to achieve their net-zero targets. Trees and other forms of green vegetation are vital to reducing these emissions as they act as carbon “sinks”, absorbing and using CO2 for growth. According to the Arbor Day Foundation, a nonprofit conservation and education organization, a mature tree absorbs CO2 at a rate of 48 pounds per year. In one year, an acre of forest can absorb twice the CO2 produced by the average car's annual mileage.

For years, conservation experts working to mitigate climate change and other environmental issues have targeted deforestation, but these efforts do not always include trees that grow outside forests. That focus excluded trees that could not only be important carbon sinks but also contribute significantly to the ecosystems and economies of the surrounding populations. With many current methods for studying trees’ carbon content only including forests, the use of AI to better identify trees outside of forests is an added tool to provide more clarity in calculating carbon content. Establishing an accurate count of trees in arid and semi-arid areas provides crucial information for researchers, policymakers, and conservationists. In addition, attaining accurate measurements of tree size and density provides vital data for improving on-the-ground conservation efforts.

The project involved an initial data collection of identifying and labeling 90,000 individual trees, with assistant professor of geography Martin Brandt, taking more than a year to code the training data. Satellite images of these manually traced trees were then fed into the tree detection framework based on U-Net — a convolutional neural network developed for biomedical image segmentation. Through supervised learning, the model was trained to recognize individual trees from the images shown, with the help of field knowledge in combination with image-interpretation skills. The results from the machine-based identification were then compared with the manually coded data and field data and showed a high level of accuracy. Using NASA and Blue Waters supercomputers, the team ran the program on the full study area and was able to identify a surprising number of trees for a region often assumed to support little vegetation.

Training AI to Better Identify Objects for Growing Use Cases

RECOMMENDATIONS


According to ABI Research’s Artificial Intelligence and Machine Learning Market Data (MD-AIML-106), the total shipments for cloud AI training chipsets is forecasted to increase to 3.8 million in 2025, from 1.4 million in 2020, growing at a 5-year CAGR of 15.1%, shows more AI training use cases developing over time. Being able to synergize AI training with commercial satellite imagery specifically for identifying trees will help provide a baseline in which the effectiveness of future efforts to revitalize and reduce deforestation can be measured. This can be further expanded beyond the dryland regions of West Africa to other arid and semi-arid areas, allowing more carbon sinks to be discovered.

Apart from training AI algorithm to identify trees, other satellite image databases can be used to train the AI to identify ships, planes, and other objects. In China, the Chinese Academy of Sciences released a fine-grained object recognition in the high-resolution remote sensing imagery (FAIR1M) database, containing detailed information of more than a million locations aimed at reducing errors made by AI when identifying objects. With advances in AI technology and the growing number of databases available to train AI, uses cases such as identifying trees to support climate change efforts will continue to improve and enable more countries to be better equipped in achieving net-zero emission targets.