Tree Species Classification on the Bill Williams National Wildlife Refuge

Submitted by Anonymous (not verified) on

Southwest river and floodplain dynamics have been dramatically altered over the last century by human activities, invasive plants, and climate change. The Bill Williams National Wildlife Refuge encompasses one of the few naturally occurring stands of cottonwood and willow riparian forest in southwestern Arizona. Sonoran Desert riparian areas play a vital role in maintaining a variety of plant and animal species; migratory and resident songbirds, waterfowl, butterflies, mammals, amphibians, and other aquatic species depend on these unique habitats, which differ strongly from surrounding upland vegetation. Since the 1990s, the U.S. Army Corps of Engineers has sought to improve native plant regeneration and establishment along the Bill Williams River by allowing “environmental flow releases” from the Alamo Lake Dam.

The FWS is collecting data and conducting analyses to monitor the impacts of these flow releases. Airborne laser altimetry (lidar) and 4-band color infrared (CIR) imagery were acquired during the fall of 2014 to determine current river channel geomorphology and vegetation structure and composition. Individual field measurements and differentially corrected UTM coordinates were collected for common tree and shrub species in December 2014 to compare spectral and structural differences. Lidar height models showed a strong relationship with shrub and tree heights measured on the ground (p ≤ 0.0001, r2 = 0.95). CIR image spectra and lidar point density metrics were used to classify individual tree and shrub species (e.g., Goodding’s Willow (Salix gooddingii), Fremont Cottonwood (Populus fremontii), mesquite (Prosopis spp.), creosote (Larrea tridentata), and salt cedar (Tamarix spp.)), but they showed low overall classification accuracy (65%) using a pixel-based random forest classifier. Lidar canopy height and point density metrics from four height strata improved discrimination among older cohorts of native tree species, but a pixel-based classifier was not able to distinguish young native trees from salt cedar and other species. Object-oriented classification methods and higher spectral resolution WorldView-3 satellite data will be used to improve class accuracy. Validated and accurate maps of vegetation composition, structure and river channel morphology will be compared with long-term pre- and post-release flow datasets to assess vegetation changes associated with hydrologic disturbance and environmental flow rates.

A 4-band CIR digital photograph and lidar-derived data layers used for discriminating individual tree species on the Bill Williams National Wildlife Refuge.

A 4-band CIR digital photograph and lidar-derived data layers used for discriminating individual tree species on the Bill Williams National Wildlife Refuge. 

Random forest-derived multidimensional scaling (MDS) plot based on lidar and image spectra showing A) tree species and predictor variables important to distinguishing individual species and B) tree species overlaid with lidar canopy height (Gaussian fit), which was the most important predictor for species classification. Older cottonwood trees were classified with 82% accuracy based primarily on canopy height. Predictor variables shown are the lidar canopy height (ldr_ht); lidar relative density at 5 to 10 m (rds5_10), 10 to 15 m (rds10_15), and 15 to 35 m height (rds15_35); normalized difference vegetation index (NDVI); and bare earth elevation (el). Individual spectral bands were not important predictors.

Random forest-derived multidimensional scaling (MDS) plot based on lidar and image spectra showing A) tree species and predictor variables important to distinguishing individual species and B) tree species overlaid with lidar canopy height (Gaussian fit), which was the most important predictor for species classification. Older cottonwood trees were classified with 82% accuracy based primarily on canopy height. Predictor variables shown are the lidar canopy height (ldr_ht); lidar relative density at 5 to 10 m (rds5_10), 10 to 15 m (rds10_15), and 15 to 35 m height (rds15_35); normalized difference vegetation index (NDVI); and bare earth elevation (el). Individual spectral bands were not important predictors.

Platform
Author Name
Steven Sesnie
Author Email
steven_sesnie@fws.gov