COS 116-3 - Plant species mapping using integrated airborne lidar and hyperspectral imagery across multiple functional groups

Thursday, August 11, 2011: 2:10 PM
13, Austin Convention Center
Kyla M. Dahlin, Climate & Global Dynamics, National Center for Atmospheric Research, Boulder, CO and Gregory P. Asner, Department of Global Ecology, Carnegie Institution for Science, Stanford, CA
Background/Question/Methods

The ability to map plant species distributions has long been one of the key goals of terrestrial remote sensing. Being able to create accurate and contiguous maps of canopy species would allow for better resource assessment and management, as well as potentially to address fundamental questions in community ecology. Recent advances in both the types of data that can be collected remotely and in available analytical tools like multiple endmember spectral mixture analysis (MESMA) are allowing for rapid improvements in this field.

In 2007 the Carnegie Airborne Observatory (CAO) acquired high resolution lidar and hyperspectral imagery of Jasper Ridge Biological Preserve (Woodside, California). The hyperspectral sensor, JPL’s Airborne Visible Infrared Imaging Spectrometer (AVIRIS) samples reflected sunlight in 10 nm increments across the 380-2510 nm range.  On this flight the AVIRIS pixel resolution was 2.7 x 2.7 m. To build a spectral library, 415 GPS points were collected in the field, made up of 44 plant species, six plant categories (for nonphotosynthetic vegetation), and four substrate types. Using the lidar data to select the most illuminated pixels as seen from the aircraft (based on canopy shape and viewing angle), we then reduced the spectral library to only the most fully lit pixels (“endmembers”). The AVIRIS data were stratified by lidar-based height into five classes. Two and three endmember models were then applied using MESMA and each pixel was assigned a species or plant category based on the highest endmember fraction.

Results/Conclusions

To validate the approach, an independent set of 200 points was collected throughout the study site and the four most dominant species within a 3 m radius of each point were identified (area equivalent to approximately four pixels in the hyperspectral imagery). The overall accuracy of the integrated lidar - MESMA approach was 92.7%, with a sensitivity of 32.3% (observed presences that were predicted), and a specificity of 96.4% (observed absences that were predicted). The most successfully mapped species included common native woody plants like shining willow (Salix lucida), chamise (Adenosotoma fasciculatum), blue oak (Quercus douglasii) and coast live oak (Quercus agrifolia), as well as the invasive yellow star-thistle (Centaurea solstitialis) and the rare serpentine grasslands. These results suggest that this integrate approach may significantly improve our ability to identify individual species and vegetation types in structurally diverse ecosystems.

Copyright © . All rights reserved.
Banner photo by Flickr user greg westfall.