Predicting forest characteristics with ALS-derived point cloud metrics and multispectral imagery in the Netherlands

By Michiel van Bemmelen

Abstract
Currently there has been limited use of Actueel Hoogtebestand Nederland/Current Dutch Elevation (AHN)-based metrics based on Airborne Laser Scanning (ALS) data in forest research. The Nederlandse Bosinvintarisatie/ National Forest Inventory (NBI) could always use geospatial help. Most forest metric research has been on small amount of forest characteristics.

The aim is to test how well a broad range of forest characteristics can be predicted using a Machine Learning (ML) model based on AHN data. Additionally, can adding Sentinel-2 data improve the accuracy, compare Random Forest Regression (RFR) and Linear Regression (LR) and does stratification on a deciduous/coniferous (DC) split improve the accuracy.

The R2 and RMSE% of the numerical variables of interest (VOIs) range from 0.24-0.65 and 63.7-19.3 and error rate for categorical VOIs range from 71.5-14.7%.
This research found that forest characteristics of the NBI can be predicted by AHN data but there are differences in accuracy between VOIs, for some VOIs the model accuracy increases by adding Sentinel-2 data. There were no differences in accuracy found between ML methods between stratified and non-stratified models.