Forest Inventory With High-Density Uav-Lidar: Machine Learning Approaches For Predicting Individual Tree Attributes

The high dimensionality of data generated by Unmanned Aerial Vehicle(UAV)-Lidar makes it difficult to use classical statistical techniques to design accurate predictive models from these data for conducting forest inventories. Machine learning techniques have the potential to solve this problem of modeling forest attributes from remotely sensed data. This work tests four different machine learning approaches namely Support Vector Regression, Random Forest, Artificial Neural Networks, and Extreme Gradient Boosting on high-density GatorEye UAV-Lidar point clouds for indirect estimation of individual tree dendrometric metrics (field derived) such as diameter at breast height, total height, and timber volume. A total of 370 trees had their dbh and height measured for validation purposes. Using LAStools we generated normalized Light Detection and Ranging (Lidar) point clouds and created a raster canopy height model at a 0.5 x 0.5 m spatial resolution following the construction of a digital terrain model and a digital surface model. The R package 'lidR' was set with the functions tree_detection (local maximum filter algorithm) and lastrees. Subsequently, we applied the function tree_metrics to extract individual metrics. Machine learning techniques were applied to the derived metrics to estimate dendrometric field measures. The machine learning models (MLM) with optimal hyperparameters showed similar predictive performances for modeling the variables diameter, height, and volume. All models had a rRMSE below 15% (for diameter at breast height), 9% (for height) and 29% (for volume). The Support Vector Regression algorithm showed the best performance. Our work demonstrates that all tested machine learning models are adequate and robust to handle the high dimensionality of UAV-Lidar data for the estimation of individual attributes, with Support Vector Regression model being the best performer in terms of minimal error rates.