Use of LiDAR and machine-learning to predict soil attributes of managed forests

An understanding of soil properties is fundamental to assessing the character of vegetation that soil will support. By relating field data on soil characteristics to topographic covariates derived from LiDAR datasets, a method was developed that uses LiDAR data to predict the soil moisture regime and soil nutrient regime. A set of topographic covariates was created from a LiDAR derived digital elevation model at multiple spatial resolutions. Random Forest, a decision tree learner, was used to predict soil moisture regime and soil nutrient regime for two case studies, the Eagle Hill forest near Savona, BC., and Eastern Nova Scotia. Tests were processed using covariates based on various filtering window sizes and different data imbalance correction methods, as ground truth data was extremely imbalanced. While variation between results was minimal, there was a trend of smaller filtering windows and less complex data balancing techniques being more effective.