Unsupervised machine learning of LiDAR-derived forest structure relevant to ladder and canopy fuels in Minnesota's Superior National Forest

Abstract

Understanding forest canopy structure is essential for monitoring ecosystem function, carbon dynamics, and wildfire risk. In this study, we developed an unsupervised classification framework that integrates 21 LiDAR-derived vertical structure metrics with Principal Component Analysis (PCA) and Self-Organizing Maps (SOM) to characterize canopy heterogeneity across the Kawishiwi Ranger District of the Superior National Forest in Minnesota, USA. The resulting structural classes support ladder- and canopy-fuel–relevant assessment and management. The PCA-based dimensionality reduction preserved key structural variation, while the SOM algorithm maintained topological relationships among feature vectors, enabling ecologically meaningful clustering. We applied K-means clustering to the trained SOM and selected the optimal number of clusters using the Silhouette score, identifying four distinct structural types—ranging from tall, multilayered canopies to sparse early-successional stands—whose spatial patterns support resilience-oriented planning, fuel-relevant structural stratification, and a district-wide baseline for monitoring and prioritizing areas for further evaluation under local objectives and constraints. Cluster-based canopy structure mapping revealed strong alignment with ecological conditions such as successional stage and disturbance history. This approach demonstrates the capacity of LiDAR and unsupervised machine learning to support scalable forest structure assessment. While promising, future integration with field-based validation and supervised learning could enhance ecological interpretability and operational applicability.

Citation

Salehnia, Nasrin; Wolter, Peter T.; Sturtevant, Brian R.; Abbas Iossifov, Dalia. 2026. Unsupervised machine learning of LiDAR-derived forest structure relevant to ladder and canopy fuels in Minnesota's Superior National Forest. Ecological Informatics. 95: 103726. 14 p. https://doi.org/10.1016/j.ecoinf.2026.103726.