Identifying alpine treeline species using high-resolution WorldView-3 multispectral imagery and convolutional neural networks

Abstract. Alpine treeline systems are remote and difficult to access, making them natural candidates for remote sensing applications. Remote sensing applications are needed at multiple scales to connect landscape-scale responses to climate warming to finer-scale spatial patterns, and finally to community processes. Reliable, high-resolution tree species identification over broad geographic areas is important for connecting patterns to underlying processes, which are driven in part by species' tolerances and interactions (e.g., facilitation). To our knowledge, we are the first to attempt tree species identification at treeline using satellite imagery. We used convolutional neural networks (CNNs) trained with high-resolution WorldView-3 multispectral and panchromatic imagery, to distinguish six tree and shrub species found at treeline in the southern Rocky Mountains: limber pine (Pinus flexilis), Engelmann spruce (Picea engelmannii), subalpine fir (Abies lasiocarpa), quaking aspen (Populus tremuloides), glandular birch (Betula glandulosa), and willow (Salix spp.). We delineated 615 polygons in the field with a Trimble geolocator, aiming to capture the high intra- and interspecies variation found at treeline. We adapted our CNN architecture to accommodate the higher-resolution panchromatic and lower-resolution multispectral imagery within the same architecture, using both datasets at their native spatial resolution. We trained four- and two-class models with aims to (1) discriminate conifers from each other and from deciduous species, and (2) to discriminate limber pine – a keystone species of conservation concern – from the other species. Our models performed moderately well, with overall accuracies of 44.1 %, 46.7 %, and 86.2 % for the six-, four-, and two-class models, respectively (as compared to random models, which could achieve 28.0 %, 35.1 %, and 80.3 %, respectively). In future work, our models may be easily adapted to perform object-based classification, which will improve these accuracies substantially and will lead to cost-effective, high-resolution tree species classification over a much wider geographic extent than can be achieved with uncrewed aerial systems (UAS), including regions that prohibit UAS, such as in National Parks in the U.S.