A tale of two "forests": random forest machine learning AIDS tropical forest carbon mapping

Area of Science:

• Ecology
• Remote Sensing
• Machine Learning

Background:

• Accurate tropical forest carbon stock mapping is crucial for climate change mitigation mechanisms like REDD+.
• Remotely-sensed data and machine learning offer potential for improved carbon mapping.
• Traditional stratification methods have limitations in spatial accuracy.

Purpose of the Study:

• To evaluate the performance of the Random Forest algorithm for upscaling airborne LiDAR-based carbon estimates.
• To compare Random Forest with and without spatial context against traditional stratification.
• To assess the suitability of Random Forest for spatially-explicit carbon stock mapping in the Western Amazon.

Main Methods:

• Airborne LiDAR data was used to estimate forest carbon stocks.
• Random Forest machine learning algorithm was applied with and without spatial coordinates (x, y).
• A large-scale validation (8 million hectares) was performed, comparing results to regional stratification.

Main Results:

• Random Forest incorporating spatial context explained 59% of carbon estimates, outperforming stratification (37%) and Random Forest without spatial context (43%).
• The inclusion of spatial context improved Root Mean Square Error (RMSE) from 33 to 26 Mg C ha(-1).
• This represents a 60% improvement in explained variation compared to models without spatial context.

Conclusions:

• Spatial context is a critical factor for improving Random Forest model performance in carbon stock mapping.
• This approach offers substantially improved carbon stock modeling for climate change mitigation.
• The findings support the use of spatially-explicit Random Forest for REDD+ and similar initiatives.