Abstract
Although the hyperspectral data with numerous bands provides abundant information for mineral identification, the accuracy of mineral mapping is largely constricted by the high correlation between adjacent bands. Band selection, a way to reduce the high dimensionality of hyperspectral data, is essential to improve the discrimination of minerals with similar spectral feartures. Random Forest (RF) algorithm performs effectively in variable selection from big data of limited observations. However, whether it’s applicable to hyperspectral data is unknown. In this study, 63 spectra of 7 mineral types are collected and a stepwise approach of RF-based band selection is applied to the spectral data with 256 bands. Optimizing input parameters including “ntree”(number of trees) and “mtry”(random selected variables), RF subsets the spectra and selects the most discriminatory bands according to the variable importance. Meanwhile, the Out of Bag (OOB) error output by RF is utilized to determine the number of bands to be remained. The plotting results show that the selected bands indeed capture the distinctive features, and the RF is proved to be as accurate but more efficient than another method N-dimensional Spectral Solid Angle (NSSA). To quantitatively assess the performance of RF, the mapping results from spectral subset was compared to the mapping with raw data. The accuracy improved by RF-based variable selection for mineral mapping demonstrates the feasibility of RF to hyperspectral band selection.