Are all those bands really necessary? Exploring data reduction and variable importance to improve statistical models for Arctic and boreal plant functional type mapping using UAV-based imaging spectroscopy.

Abstract:

Measuring terrestrial biodiversity and vegetation composition and structure depends on accurately mapping plants at the finest taxonomic resolution feasible. Arctic and boreal vegetation can be a mosaic of small patches of intermixed vegetation, which is a challenge for remote sensing and mapping of plants. Imaging spectrometers mounted on unoccupied aerial vehicles (UAV) provide large amount of data about a target’s reflectance properties while also generating data with small enough pixels to resolve small vegetation patches. However, these sensors with hundreds of bands are more difficult to operate and produce a cumbersome amount of data, possibly more than is necessary for the tasks of mapping plants at fine taxonomic scale. We experimented with reducing the amount of data from a UAV-based VNIR imaging spectrometer to answer the question “What is the optimal spectral information for mapping plants in the ABoVE domain?” and “What areas of the electromagnetic spectrum are most important for this task?”.

We built statistical models that classified pixels of UAV-based imaging spectrometer data. We iteratively and systematically reduced the spectral data used in the models, observed changes in model fit and predictor importance. Our training data was composed of 1000 ground-based reflectance measurements of 100 different vascular and non-vascular plants. Models built on these spectra were applied to the imaging spectrometer data and validated with ground measurements. Our results used 19 different plant functional types (PFTs) as the response variables. We found that removing intercorrelated predictors (r=>0.95) and retaining the 80 most important predictors (out of over 200) maintained model fit compared to a saturated model. Carter vegetation indices were the single most important predictor of PFT classes followed by other vegetation indices. Narrow band (5nm) reflectance features at 412 and 467 nm were the fifth and sixth most important variables. While the overall model error was <30%, broadleaved shrub categories were difficult to separate but conifer trees and lichen classes were successfully separated. These results suggest that while some data reduction is possible, eliminating enough bands to approach a multispectral sensors bandpass results in far poorer ability to discriminate PFTs.