A Machine Learning Framework for Identifying Shrubs vs. Trees and Live vs. Dead in the Boreal Forest of Interior Alaska

Abstract:

Fire is a dominant forest disturbance process in the boreal forests of interior Alaska and has a direct impact on increased carbon emissions and reduction of above ground biomass. In boreal environments, both trees and shrubs play important roles in regional-scale ecosystem dynamics including nutrient cycling, forest succession, and fire behavior. Given the vast extent of this region, most previous vegetation mapping efforts have utilized course- to moderate- resolution satellite data, there has yet to be a spatially extensive and high-resolution map for Interior Alaska that includes the composition of trees versus shrubs and live versus dead status. This would reduce the uncertainty in biomass and carbon estimation, generate baseline biogeographical characterizations of ecosystem structure and composition, and benchmark vegetation conditions across a landscape where trees and shrubs may respond differently to rapid warming, in terms of competition, growth rates, and susceptibility to drought, disease and insects. We present a framework to generate high resolution maps of trees vs shrubs and live vs dead based on the combination of forest inventory plots and G-LiHT data using a hybrid object- and pixel-based approach with convolutional neural network (CNN). Specifically, we describe a framework for the multi-label problem of classifying forest and shrub and among them live vs. dead status. The training, validation, and test datasets are composed of lidar, very fine resolution hyperspectral (1m) and stereo imagery (3 cm). In this framework we also study the contribution of the different data modalities to CNN classifier accuracy identifying forest and shrub and live vs. dead classes. Fine resolution trees and shrub live vs. dead classification maps will help in understanding of the fine scale processes related to canopy and surface layer consumption by fire and drought. Importantly, including shrubs biomass will improve the carbon monitoring system for Interior Alaska.