Applying Convolutional Neural Network (CNN) Classification of WorldView-3 Satellite Imagery to Distinguish Vegetation Species at Treeline in Rocky Mountain National Park

Treeline systems are influential for the retention and distribution of snowpack on the landscape, affecting downstream water uses. With increasing global average temperatures, high-elevation trees and shrubs are expected to advance upslope, though the rate and pattern of this change is species-dependent. We aim to use both multispectral (1.2 m) and panchromatic (0.32 m) WorldView-3 satellite imagery to distinguish dominant tree and shrub species in treeline communities in Rocky Mountain National Park to inform predictions of treeline movement. We collected 1000 perimeters of 6 tree and shrub species using a Trimble Geo7X: limber pine (Pinus flexilis), Engelmann spruce (Picea engelmannii), subalpine fir (Abies lasiocarpa), glandular birch (Betula glandulosa), willow (Salix species: glauca, brachyphora, and hybrids), and aspen (Populus tremuloides). We spatially separated this dataset into training, validation, and testing datasets—all data fell within a single, orthorectified and atmospherically corrected (ACOMP method), 25 km2 WV-3 image. We manually selected pixels that fell entirely within the training sample perimeters, visually correcting for spatial inaccuracies by referencing the high-resolution imagery. We trained a convolutional neural net (CNN) model, using both multispectral and panchromatic WV-3 image patches: a 40-by-40 m box centered on each ground-truth pixel. The CNN draws spectral, textural, and spatial information from the image patches for pixel-based image classification. The class frequencies (number of pixels) were: 0.229 subalpine fir; 0.157 glandular birch; 0.166 Engelmann spruce; 0.220 limber pine; 0.038 aspen; 0.190 willow. We evaluated the model using 5-fold cross-validation. The CNN model consistently outperformed a trivial model, with a top-1 accuracy of 0.420 (compared with 0.229 for the trivial model), a top-2 accuracy of 0.665 (compared with 0.449), and a top-3 accuracy of 0.798 (compared with 0.639). Future work will incorporate environmental predictors—including a digital elevation model, the climatological frequency of late-season snow cover at each pixel over several decades, soil type, and vegetation indices such as NDVI—to improve classification accuracy.