Advancing forest carbon stocks' mapping using a hierarchical approach with machine learning and satellite imagery

Area of Science:

• Ecology
• Forestry
• Remote Sensing
• Machine Learning

Background:

• Forest monitoring is crucial for biodiversity, resource management, and climate change mitigation.
• Accurate forest carbon stock estimation requires detailed data on species, age, height, and basal area.
• Remote sensing and machine learning offer scalable solutions for analyzing large forest areas.

Purpose of the Study:

• To develop an automated machine learning pipeline for forest carbon stock estimation.
• To utilize freely available Sentinel-2 satellite observations for predicting forest structure and carbon stocks.
• To provide an end-to-end solution for large-scale forest carbon assessments.

Main Methods:

• Employed multispectral Sentinel-2 remote sensing data.
• Utilized the Extreme Gradient Boosting (XGBoost) algorithm for classification and regression.
• Integrated management-level inventory data as reference measurements.
• Explored direct and hierarchical approaches for carbon stock estimation.

Main Results:

• Achieved an F1-score of 0.75 for species prediction.
• Obtained R-squared values of 0.75 for stand age, 0.58 for height, and 0.56 for basal area.
• Successfully estimated stem carbon stock using remote sensing data and conversion factors.

Conclusions:

• The developed pipeline offers a robust and accurate method for large-scale forest carbon stock estimation.
• The approach effectively integrates satellite imagery, machine learning, and inventory data.
• This method enables more reliable global assessments, even with limited annotated datasets.