Linear Spectral Mixture Analysis Based Approach with Regard to Neighboring Pixels for Multispectral Image Classification

Nareenart Raksuntorn

Faculty of Industrial Technology, Suan Sunandha Rajabhat University, Thailand

E-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.

andWinai Raksuntorn

Faculty of Engineering, Thammasat University, Thailand

E-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.

Manuscript received October 17, 2015

Revised November 16, 2015


This paper presents an algorithm to implement a linear spectral mixture analysis (LSMA) based approach with regard to neighboring pixels for multispectral image classification. This approach allows the implementation of LSMA in the way of the traditional manner does, i.e., an unconstrained LSMA (ULSMA). In our approach, a given pixel and its neighboring pixels are involved in the unmixing process. A given pixel will be classified as a material corresponding to the maximum abundance if the resulting abundances are all non-negative. Otherwise, the algorithm goes to a searching process. In this process, the algorithm searches for abundances obtained from a given pixel and its neighboring pixels that are satisfied three conditions: non-negativity, sum-to-one, and minimum mean squares error on reflectance reconstruction (MSER). Here, the endmember sets used in the searching step are varied where endmembers corresponding to nonnegative numbers are fixed and other endmembers can be varied. Five experiments for a real multispectral image are conducted including an ULSMA, a multiple endmember spectral mixture analysis (MESMA), the proposed approach with one-pixel and two-pixel distances, and a maximum likelihood (ML) estimation. The first three methods are based on the LSMA method while the last one, ML, is based on parameter estimation. Three major material types: water, vegetation, and road/building, are present in the image scene. The experimental results demonstrated that the proposed method offered higher classification accuracy when compared to the ULSMA and MESMA. The overall accuracies were about 89.50% and 88.50% for one-pixel and two-pixel distances, respectively. However, when compared to the ML classifier, the proposed approach provided a lower accuracy in overall. Specifically, the proposed approach yielded higher accuracies for two classes: water and vegetation. But for the road/building class, the performance of the proposed approach was lower due to endmember variability. When compare the computational cost, the proposed approach outperformed both MESMA and ML. 

Keywords: Multispectral image; classification; linear spectral mixture analysis

pdf File Size: 365.12 KB

Mahanakorn University of Technology

140 Moo 1, Cheum-Sampan Road, Nongchok, Bangkok, Thailand 10530

Tel: +(662)988-3655  Fax: +(662)988-4027

designed by