COS 21-8 - Mapping vegetation biodiversity of semi-arid ecosystems using hyperspectral remote sensing

Monday, August 7, 2017: 4:00 PM
B114, Oregon Convention Center
Hamid Dashti1, Nancy F. Glenn1, Nayani Illangakoon1, Susan Ustin2, Marie-Anne de Graaff3, Lejo Flores1, Yi Qi4 and Lucas Spaete5, (1)Geosciences, Boise State University, Boise, ID, (2)Land, Air and Water Resources, UC Davis, Center for Spatial Technologies and Remote Sensing, Davis, CA, (3)Department of Biological Sciences, Boise State University, Boise, ID, (4)Department of Land, Air and Water Resources, University of California, Davis, Davis, CA, (5)Geosceinces, Boise State University, Boise, ID

Vegetation biophysical and chemical parameters such as Leaf Area Index (LAI), Nitrogen (N) content and vegetation cover are important indicators of ecosystem biodiversity. Optical hyperspectral (400-2500 nm) remote sensing has been widely used for ecosystem biodiversity mapping. Mapping biodiversity in non-homogeneous systems such as semi-arid ecosystems using common remote sensing techniques is fraught with challenges. Sparse plant distribution and low LAI of vegetation in these ecosystems intensify the influence of background soil on the vegetation reflectance (the mixed pixel effect). The main objective in this study is to improve both mapping of LAI and N using hyperspectral imaging at different spatial scales, and mapping shrub cover at sub-pixel resolutions. We aim to answer, how much uncertainty is introduced in the parameter estimation when scaling up from leaf to landscape? and; how can spectral unmixing techniques help us map vegetation cover at subpixel resolutions? Our study is focused on the shrublands within the Great Basin (GB), western US. The GB is experiencing drought, wildfire, and species invasion which collectively are changing the structure and function of the ecosystem.


During 2014 and 2015, airborne hyperspectral data (AVIRIS-NG) and extensive field data including biophysical and biochemical variables (e.g. LAI, N, carbon) and spectroscopy measurements were collected across an elevation and precipitation gradient in the GB. We are using statistical methods such as Partial Least Squares Regression (PLSR) and Least Absolute Shrinkage and Selection Operator (LASSO) analysis to predict shrub parameters from the hyperspectral data. Our preliminary results show promising performance of statistical predictions (0.30< R-squared <0.83) for various variables at different scales. However, the transferability of statistical models to different sites, sensors and/or time remains questionable. Physically-based Radiative Transfer Models (RTMs) provide more stable predictions. We will test 1-D (PROSAIL) and 3-D (DART) RTMs for estimating the same parameters we are estimating with statistical techniques. We will compare the RTM and statistical predictions and validate both with ground truth data. To map shrub cover at sub-pixel levels, we used multiple endmember spectral mixture analysis coupled with spectral endmember bundling. Preliminary results show correlations between hyperspectral derived shrub cover with field (R2 = 0.54), and lidar (R2 = 0.88) derived covers where shrub cover is >25%. The overall results of this study will provide a framework for ecologists to map semi-arid ecosystem vegetation biodiversity from local to global scales.