Scientists teach AI to assess forest carbon stocks from satellite images

Scientists from Skoltech, together with colleagues from Irkutsk National Research Technical University and the AIRI Institute of Artificial Intelligence, have developed a machine learning algorithm that determines key forest characteristics and estimates the amount of carbon accumulated in them from satellite images.

This was reported by Skoltech's press service.

How the system works

The algorithm was trained on data from forest management records, Sentinel-2 satellite images, and topographic maps of forests in the Sakhalin region. It is capable of determining:

Dominant tree species Stand age Tree height Timber stocks Amount of accumulated carbon

An important advantage of the new model is its built-in uncertainty assessment of the prediction. The algorithm does not output a single value but a range, along with an indication of confidence level. In complex and heterogeneous forest areas, the margin of error automatically increases, and the system explicitly shows this.

Results

The XGBoost algorithm demonstrated the best performance:

Accuracy in determining tree species: 83% Accuracy in determining age: about 70% Accuracy in estimating timber and carbon stocks: 53–63%

Why this is needed

Forests play a key role in absorbing carbon dioxide and combating climate change. Traditional ground-based surveys require enormous resources, while the new technology enables rapid and relatively low-cost monitoring of large areas.

"The developed tool combines satellite data with uncertainty estimation algorithms for rapid forecasting of forest characteristics," noted Alexander Bernstein, a professor at the Skoltech Center for Artificial Intelligence.

In the future, the scientists plan to scale the model and adapt it for other regions of Russia and the world. The technology could be useful for forestry, climate research, and carbon monitoring.

In brief

Scientists at Skoltech have created an AI algorithm that assesses forest characteristics and carbon stocks from satellite images. The model not only provides a prediction but also shows its level of reliability. XGBoost achieved the best results. The development will make it possible to significantly speed up and reduce the cost of forest monitoring and assessing their contribution to climate stability.

Follow NEWS.am Tech on Facebook and Twitter