From point to pixel: Using in situ measurements to validate and derive higher level NEON hyperspectral data products
Understanding the drivers of ecological change through time requires detailed, multi-dimensional information that quantifies functional and structural properties of ecosystems, from points to pixels on, below and above the ground and from local to broad, continental scales. NEON hyperspectral imagery (HSI), coupled with lidar, and extensive biological, chemical and physical ground measurements, collected across the United States, over a 30 year time period will provide a unique opportunity to understand ecological change across critical spatial and temporal scales. However, NEON HSI data is collected at very fine spectral (~5nm) and spatial (~1m) resolution. As such, the pixel size from the NEON HSI instrument is generally smaller than most vegetation canopies, including trees and shrubs. It also provides much greater spectral fidelity than other sensors. For example, in the Near Infrared Region, the NEON HSI instrument provides 3 bands corresponding to Landsat 7’s single NIR band. These differences in instrument resolution require NEON to revisit methods used to derive commonly used spectral vegetation indices (SVIs) like NDVI and EVI.
One primary use of this new airborne HSI dataset is the creation of a Canopy nitrogen (N, %) product that estimates %N across an unprecedented range of ecosystem and vegetation types (at NEON field sites). In this study, we explore various published methods to estimate Canopy N including typical spectral vegetation indices (SVIs, e.g. NDNI), partial least squares regression (PLSR) and using wavelengths that have been found to be strongly correlated to %N (eg near IR region bands) as applied to NEON HSI imagery. We validated published approaches for calculating N as compared to in situ leaf level measured N and scaled these estimates to canopy level values to determine a best approach.
Initial results suggest that the typical band math / SVI approach may not be appropriate to estimate canopy N across a diverse array of sites. One the other hand, PLSR,or other methods may be more robust across multiple sites with varying degrees of canopy diversity and complexity. Further, it may be necessary test these relationships by to isolating pixels that best represent top of the canopy conditions (i.e., sunlit crown pixels) as seen by the HSI instrument using a fusion approach that includes lidar (light detection and ranging) and NDVI values extracted from the HSI.