Spectral Band Subsetting for the Accurate Mining of Two Target Classes from the Remotely Sensed Hyperspectral Big Data

Abstract

Due to the outstanding topical development of sensor technology, there is a substantial increase in the spatial, spectral, and temporal resolution of the remotely sensed hyperspectral data. The increase in the resolution of the data in turn has increased the volume, velocity, and variety of the data has contributed to identify the hyperspectral data as ‘Big Data.’ On the one hand, hyperspectral big data is a rich source of information for several applications as it consists of varieties of classes that include natural and man-made land covers. On the other hand, mining of the required class by the application from such a massive data turns out to be very difficult and hence requires a smart hyperspectral data analysis. Realizing that the user could be interested to mine just one or two target classes from among several classes as required by the given application, we propose an effective technique to handle this voluminous data by focusing on just one target class at a time. This research contribution includes the designing of a spectral band subsetting to address the dimensionality reduction problem by focusing on just one target class in parallel with the second target class. The proposed spectral subsetting is carried out in two stages. In the first stage, the most significant spectral band of both the target classes is instigated independently, and in the second stage, they are merged and validated. As there is a possibility that two target classes are overlapping with each other due to their spectral similarities, a method is also proposed to solve the overlapping of the target classes. The experiment is carried out on a benchmark data set, namely AVIRIS Indiana pine, ROSIS Pavia University.