Leveraging ICESat-2 and Machine Learning for High-resolution Canopy Height Mapping in Boreal Forests

Keywords: ICESat-2, Forest canopy height, PlanetScope, ArcticDEM, Machine learning, Random forest, Boreal forest
Abstract Type: Paper Abstract

Abstract

Forest canopy height is a critical geophysical parameter for assessing forest degradation-restoration, modeling ecosystems, and estimating carbon content. An accurate estimation of forest canopy structure is crucial to understanding the current and future global carbon cycle dynamics due to its strong association with aboveground biomass, primary productivity, and forest ecosystem services. This study presents a novel approach for generating a high-resolution (5-m) canopy height model by integrating cloud-optimized ICESat-2 vegetation height data processed at 5-m with very high-resolution multispectral and topographic imagery. A machine learning workflow was developed using ICESat-2-derived vegetation height combined with spectral, textural, and topographic features from PlanetScope and ArcticDEM. The performance between three machine learning models – random forest (RF), artificial neural network (ANN), & gradient boosting machine (GBM) were evaluated. Results indicate that both RF and GBM models achieved satisfactory accuracy in predicting boreal forest canopy height, with RF proving to be the more accurate regressor. The accuracy comparisons of predicted canopy height map with multiple reference datasets indicate that the proposed approach outperforms existing canopy height datasets by capturing finer-scale structural variations in boreal forests. The study underscores the potential of ICESat-2 canopy height data to enhance fine-scale canopy height mapping, aiding in accurate forest biomass and carbon stock assessments.