Researchers use training models to map planted and natural forests from satellite imagery

(a) World, (b) Europe, (c) Asia, (d) North America, (e) Oceania, (f) South America, (g) Africa. Photo credit: Xiao Yuelong, Tongji University.
Reforestation may seem like a simple way to combat climate change, but planted forests often encroach on natural forests, wetlands and grasslands, which can reduce biodiversity, disrupt natural habitats and disrupt carbon and water cycles.
Forest cover is increasing worldwide, but it’s difficult to know whether this is due to the regeneration and growth of natural forests or the planting of new trees. Accurately mapping these forests using remote sensing techniques could help.
However, although planted and natural forests can be distinguished on satellite images based on their characteristics, there are no comprehensive planted and natural forest maps.
A study published August 21 in the Journal of Remote Sensing presents an innovative approach to automatically generate training samples so that natural and planted forests can be accurately mapped at a spatial resolution of 30 meters.
“Accurately mapping the global distribution of natural and planted forests at high spatial resolution is difficult, but is crucial for understanding and mitigating environmental problems such as carbon sequestration and biodiversity loss,” said Xiao Yuelong, a doctoral student at the School of Surveying and Geoinformation at Tongji University in Shanghai, China.
“Traditional methods often lack sufficient training samples, which hinders the accuracy and resolution of global forest maps. Our study presents a novel approach that overcomes this limitation by generating extensive training samples through time-series analysis of Landsat imagery.”
The researchers obtained data from several different mapping systems, with the primary sources being Google Earth Engine’s Landsat imagery from 1985 to 2021, preprocessed by the United States Geological Survey, and imagery from the Sentinel-1 satellite in 2021.
The researchers also used data from the European Space Agency’s Land Cover Atlas 2021, called WorldCover2021, and the ALOS Global Digital Land Surface Model. To get around computing limitations, the researchers divided the Earth into small tiles, creating 57,559 tiles and 70 million training samples that cover the entire Earth.
To distinguish between existing natural and planted forests, the researchers used a value called disturbance frequency. Natural forests are more stable and less likely to change size based on external factors.
In comparison, planted forests are more likely to be disturbed by reforestation, deforestation, and other natural and human-induced changes. Satellite imagery can help track how often forest areas are disturbed, making it possible to distinguish between natural and planted forests.
Planted forests were considered as pixels with a disturbance frequency greater than 3. This value was calculated based on the number of disturbance events, such as afforestation events, and the reliability of the training samples. Natural forests had no disturbance events. The researchers also took into account the fact that all the images were from 1985 onwards.
The researchers used other features and characteristics to distinguish between natural and planted forests to account for plantations that may be older than 1985. Finally, to determine the accuracy of their training model, the researchers compared their maps of natural and planted forests to other studies.
This study demonstrated the feasibility of a low-effort mapping method to distinguish natural and planted forests using automatically generated training samples.
“This method is reliable for accurately mapping natural and planted forests around the world at 30-meter resolution, and the maps and training samples generated will be a valuable resource for future research and environmental management, contributing to efforts to combat climate change,” Xiao said.
Going forward, the researchers hope to incorporate improvements into the mapping system.
“We will then use the generated training samples and method mapping to regularly update and refine our global map of natural and planted forests. Our ultimate goal is to increase the accuracy and resolution of forest maps around the world, providing important data for policymakers and researchers,” Xiao said.
Other contributors include Qunming Wang of Tongji University in Shanghai, China, and Hankii K. Zhang of South Dakota State University in Brookings, South Dakota.
Further information: Yuelong Xiao et al., “Mapping the world’s natural forests and plantations at 30 m high spatial resolution,” Journal of Remote Sensing (2024). DOI: 10.34133/remotesensing.0204
Courtesy of the Journal of Remote Sensing
Citation: Researchers use training models to map plantations and natural forests from satellite imagery (September 16, 2024) Retrieved September 16, 2024 from https://phys.org/news/2024-09-natural-forests-satellite-image.html
This document is subject to copyright. It may not be reproduced without written permission, except for fair dealing for the purposes of personal study or research. The content is provided for informational purposes only.