Estimating grassland Biomass using SVM band shaving of hyperspectral data

In this paper, the potential of a band shaving algorithm based on support vector machines (SVM) applied to hyperspectral data for estimating biomass within grasslands is studied. Field spectrometer data and biomass measurements were collected from a homogeneously managed grassland field. The svm band shaving technique was compared with a partial least squares (PLs) and a stepwise forward selection analysis. Using their results, a range of vegetation indices was used as predictors for grassland biomass. Results from the band shaving showed that one band in the near-infrared region from 859 to 1,006 nm and one in the red-edge region from 668 to 776 nm used in the weighted difference vegetation index (WDVI) had the best predictive power, explaining 61 percent of grassland biomass variation. Indices based on short-wave infrared bonds performed worse. Results could subsequently be applied to larger spatial extents using a high-resolution airborne digital camera (for example, Vexcel's UltraCam (TM)).