Mutual Information-Based Hierarchical Band Selection Approach for Hyperspectral Images
Hyperspectral images consist of hundreds of spectral bands with relatively narrow bandwidth and hence records detailed information of the objects. Due to this detailed and enormous amount of information content, the use of hyperspectral images has become
- PDF / 244,551 Bytes
- 9 Pages / 439.37 x 666.142 pts Page_size
- 101 Downloads / 265 Views
Abstract Hyperspectral images consist of hundreds of spectral bands with relatively narrow bandwidth and hence records detailed information of the objects. Due to this detailed and enormous amount of information content, the use of hyperspectral images has become very popular in various fields such as land cover monitoring, agriculture, defense, etc. However, this increased spectral dimension results in increased computational complexity. Hence, the selection of minimal subset of spectral bands to represent the actual information effectively without much degradation is a challenge in the field of hyperspectral image analysis. This paper proposes a hierarchical band selection approach by constructing a spectral partition tree-based on mutual information. Initially, each spectral band has been considered as a leaf node. To minimize the redundancy of information carried by neighboring bands, in every level, new nodes are created by merging adjacent bands or group of bands, for which mutual information has been used as the deciding criterion. Finally from each group of bands, a representative band is selected which jointly form the set of selected bands. Experiment is carried out on the AVIRIS Indian Pines dataset by designing training and testing samples containing only the selected set of bands. The experimental results of the proposed method are found to be very promising and competitive with the existing techniques.
Keywords: Hyperspectral images Spectral partition tree Mutual information Support vector machine
Entropy
S. Sarmah (&) S.K. Kalita Department of Computer Science, Gauhati University, Guwahati, India e-mail: [email protected] S.K. Kalita e-mail: [email protected] © Springer Nature Singapore Pte Ltd. 2018 A. Kalam et al. (eds.), Advances in Electronics, Communication and Computing, Lecture Notes in Electrical Engineering 443, https://doi.org/10.1007/978-981-10-4765-7_78
755
756
S. Sarmah and S.K. Kalita
1 Introduction Hyperspectral sensors capture images with narrow and contiguous spectral bands covering spectrum not only from the visible range, but also from the ultra violet and infra red region. The resultant datasets are three-dimensional which are represented as data cubes of size P Q N, where P and Q are the spatial dimensions and N is the spectral dimension. Each spatial plane can be viewed as collection of two-dimensional scenes containing P Q pixels, each taken at a specific wavelength k and each pixel can be viewed as a vector consisting of N reflectance values. One such sensor is the airborne visible/infrared imaging spectrometer (AVIRIS) that captures images with up to 224 spectral bands ranging from 400 to 2500 nm [1]. With such high spectral dimension, much more detailed and discriminative information of the objects can be acquired which results in increased classification accuracy. However, performance of many supervised classification methods get strongly affected by increased dimensionality. Hence, spectral dimension reduction is a crucial step in hyperspect
Data Loading...
