Preprocessing Pipeline

Input raw data is processed using the PDAL (Point Data Abstraction Library) framework to standardize the point cloud for analysis. The pipeline begins with the removal of outliers and noise points based on statistical distance metrics. Subsequently, a ground segmentation algorithm separates terrain points from vegetation and objects.

The final step involves normalizing the point cloud to Height Above Ground (HAG) using Nearest Neighbor interpolation and applying a semantic classification. The XGBoost machine learning algorithm is deployed for point-based semantic segmentation to distinguish between vegetation, buildings, and ground classes.

Predictor Calculation

Structural metrics are calculated on a per-plot basis using a circular area of 500 m². These calculations are performed for each cell of the standardized 10m or 20m output raster grid.

Computed Metrics include:

• Z-Statistics: Mean, Standard Deviation, Skewness, Kurtosis.
• Height Quartiles (Q): Vertical distribution metrics describing canopy height levels (e.g., Q90).
• Penetrability Metrics (P): Describing signal penetration through the canopy structure (e.g., P90).
• Canopy Cover: Calculated as the ratio of vegetation returns (> 1.5m) to total signal returns.

Machine Learning Estimation

Above-Ground Biomass (AGB) prediction utilises a Stacking Regressor ensemble technique. A set of base learners, including Random Forest Regressor, Gradient Boosting (GBM), and Support Vector Regressor (SVR), generate initial predictions. These are then aggregated by a linear Meta-Learner (Ridge Regression) to minimize variance and improve generalization.

The models were trained on synthetic ALS data generated from real tree point clouds using Helios++. These inputs were acquired via TLS, and biomass was calculated for each individual tree.

The model outputs the estimated dry biomass weight per hectare (t/ha).

Project Aim

ABiLAS is a web-based platform for automated forest biomass estimation. It allows users to upload airborne laser scanning (ALS) data (LAZ format), which are processed using a pipeline of PDAL algorithms and Machine Learning models. The application visualizes the results on an interactive map, providing detailed spatial statistics on Above-Ground Biomass (AGB), vegetation height, and volume.

Partners

The project is realized by a research team from the Global Change Research Institute of the Czech Academy of Sciences (CzechGlobe) in collaboration with Masaryk University (MUNI) and CESNET, which provides the architectural support for the EnviLAB platform.