Background
The high dimensionality of imaging spectroscopy greatly improves the ability of hyperspectral mineral mapping. Hyperspectral data, consisted of continuous and numeric bands, does not only provide abundant information for mineral identification but also increase the colinearity among bands, which is obstructive to discriminate minerals of tiny difference. To make effective use of hyperspectral data it’s very critical to reduce the data dimensionality for accurate mineral mapping. Selecting the most useful bands which maximize the contrast of spectral endmembers for mapping would effectively reduce the Figure 1. Hyperspectral sensing
data dimensionality, as demonstrated in the methods of Mutual Information and Clonal selection (Jie Feng, 2014) and Indicator Kriging (Freek van der Meer, 2005). Random forest (RF), a popular statistical method based on bootstrap classification and regression tree (CART), can generate a score of variable importance using backwards variable elimination and is potential to select a small sets of non-redundant variables from the highly-correlated hyperspectral bands. However, the performance of variable selection by random forest on hyperspectral data is suspected, and how to measure the importance of correlated bands by RF is unclear either.
data dimensionality, as demonstrated in the methods of Mutual Information and Clonal selection (Jie Feng, 2014) and Indicator Kriging (Freek van der Meer, 2005). Random forest (RF), a popular statistical method based on bootstrap classification and regression tree (CART), can generate a score of variable importance using backwards variable elimination and is potential to select a small sets of non-redundant variables from the highly-correlated hyperspectral bands. However, the performance of variable selection by random forest on hyperspectral data is suspected, and how to measure the importance of correlated bands by RF is unclear either.
Objectives
Therefore, stepwise experiments are performed to discover whether this robust method of variable selection, random forest, is feasible for hyperspectral band selection. Problems are focus on: will bands selected by RF represent the key features of the minerals, what's the advantages of RF over other methods, and how much degree can the band selection improve the accuracy of mineral mapping. Through this study, we expected to obtain several band subsets with importance, the proper number of bands, performance of mineral mapping after band selection, and a comparison with other method.
Expected results
The study expects to subset the 256 spectral bands, extract endmembers with diagnostic features. Band selection attempts to improve the efficiency of mineral mapping, since useful bands are utilized to identify minerals and the rest could be removed if they are redundant or meaningless for mapping. The feasibility of RF to dimensional reduction of hyperspectral data is analyzed through error validations.
Figure 2. Expected results