Integrating UAS and Satellite Data to Improve Classification of Plant Functional Types in a Forest Landscape

Abstract

Remotely sensed land-use/land-cover (LULC) data products are an important tool for understanding landscape processes at all scales. The types of inference researchers can derive from remote sensing data depend on the quality and characterization of these LULC products. Currently, gaps in the detail of widely available vegetation cover products limit inference about processes at the level of plant functional type (PFT). Advanced remote sensing tools, machine learning, and data fusion can be employed to contribute novel methods to produce detailed, accurate vegetation cover products. Here, we combine high resolution Unoccupied Aerial Systems (UAS) multispectral, structural and textural data with large-scale satellite topographical data to classify PFTs in a forested landscape in the southern Rocky Mountains. First, we explore the relationships between topographical factors and spectral reflectance of vegetation. Second, we test if these multi-source, multi-scale data enable us to differentiate between PFTs, including woody vegetation types (evergreen trees, deciduous trees, and woody shrubs) as well as herbaceous and dead vegetation. We found varied but significant effects on spectral reflectance through interactions between PFTs and topographical variables. Northness was identified as an especially influential variable on spectral reflectance across PFTs. The inclusion of topographic features in our classification analysis improved overall classification accuracy by approximately 2%. Classification accuracies of PFTs ranged from 75% (deciduous trees) to 96% (evergreen trees). This analysis highlights the potential for improved understanding of vegetation cover on the landscape through integration of multi-source, multi-scale remote sensing data.