A comparison of spectral mixture analysis and ten vegetation indices for estimating boreal forest biophysical information from airborne data

Spectral Mixture Analysis (SMA) provides the capability to derive the percentage of sunlit crowns, background, and shadows within a remote sensing image pixel. This sub pixel scale information has been shown to consistently provide significantly improved estimates of forest biophysical information such as biomass, leaf area index (LAI) and net primary productivity (LAPP) compared to that provided by the normalized difference vegetation index (NDVI) using airborne and satellite imagery: However, a number of vegetation indices (VIs) have been proposed as an improvement to ND VI. In this paper, ten different VIs were used to predict forest biophysical parameters, and then compared with results obtained from SAM using airborne multispectral data from the NASA COVER Project, Superior National Forest, Minnesota USA. This data set was acquired over a range of solar zenith angles at the spatial resolution and spectral bands of the Landsat Thematic Mapper sensor. The following vegetation indices were derived from the remotely sensed data: NDVI, SR, MSR, RDVI, WDVI, GEMI, NLI, and three different soil-adjusted vegetation indices (SAVI, SAVI-1, SAVI-2). Results were obtained at solar zenith angles of 30 degrees, 45 degrees and 60 degrees for biomass, LAI NPP, DBH, stem density, and basal fraction. In all cases, SMA shadow fraction provided significantly better results than any vegetation indices, with improvements on the order of 20% compared to the best vegetation index results. In most cases, one or more of the new vegetation indices provided a small to moderate improvement compared to NDVI, with WDVI and SAVI-1 performing best amongst the [VIs, likely due to the inclusion of background reflectance. It is concluded that while different VIs can provide some improvements over ND VI, these appear to be functionally equivalent and fundamentally similar in a first-order sense in terms of their ability to predict forest biophysical parameters, which is surpassed by approaches based on subpixel image decomposition such as spectral mixture analysis and canopy optical reflectance modelling.