Total and component forest aboveground biomass inversion via LiDAR-derived features and machine learning algorithms

Area of Science:

• Forestry
• Remote Sensing
• Ecology

Background:

• Forest aboveground biomass (AGB) is crucial for assessing ecosystem health, productivity, and carbon stocks.
• Light Detection and Ranging (LiDAR) technology excels at capturing forest vertical structure and vegetation spatial distribution.

Purpose of the Study:

• To estimate forest total and component AGB using features derived from airborne LiDAR point cloud data.
• To rank and optimize LiDAR-derived features for AGB estimation using partial least squares regression.

Main Methods:

• Extracted 56 features from airborne LiDAR point cloud data.
• Utilized partial least squares regression for feature importance ranking and optimization.
• Employed random forest (RF) and support vector regression (SVR) algorithms for AGB estimation, validated by leave-one-out cross-validation (LOOCV) and cross-validation.

Main Results:

• Key features for AGB estimation included cumulative height percentiles (AIH), height percentiles (H), and height-related variables (Hmean, Hsqrt, Hmad, Hcurt).
• The RF algorithm achieved the best performance for total AGB estimation (R² = 0.75, RMSE = 22.93 Mg/ha via LOOCV).
• Accuracy for component AGB estimation followed the order: stem > bark > branch > leaf, with RF outperforming SVR.

Conclusions:

• LiDAR-derived features are sensitive indicators for forest total and component AGBs.
• RF algorithm provides more accurate AGB estimations than SVR, with minimal validation differences between LOOCV and cross-validation (<5%).
• Retrieval of leaf component AGB showed lower accuracy compared to other biomass components.