Large-scale probabilistic identification of boreal peatlands using Google Earth Engine, open-access satellite data,...

Area of Science:

• Earth Observation
• Machine Learning
• Wetland Science

Background:

• Peatlands are vital for biodiversity, climate regulation, and ecosystem services like carbon storage.
• Accurate peatland mapping is essential for conservation and resource management.
• Advancements in satellite data and cloud computing enable large-scale, high-resolution landcover mapping.

Purpose of the Study:

• To develop and apply a data-driven framework for mapping peatlands in Alberta, Canada.
• To utilize freely-available satellite data and Google Earth Engine for high-resolution peatland identification.
• To assess the accuracy of peatland mapping models.

Main Methods:

• Compiled extensive Earth observation datasets (radar, optical, LiDAR).
• Developed a machine learning model to predict peatland occurrence.
• Validated the model using a dedicated dataset, achieving 87% accuracy for peatland identification.

Main Results:

• The peatland occurrence model achieved 87% accuracy and a kappa statistic of 0.57.
• Distinguishing peatlands from mineral wetlands yielded 69% accuracy and a kappa statistic of 0.37.
• The model successfully mapped peatlands across a 397,958 km² area of Alberta at 10 m resolution.

Conclusions:

• A data-driven framework using satellite data and machine learning is effective for large-scale peatland mapping.
• This methodology supports accurate wetland and landcover inventories for resource management.
• The approach is scalable to geopolitical levels for environmental monitoring.