AUTOMATED CLASSIFICATION OF LAND COVER USING LANDSAT 8 OLI SURFACE REFLECTANCE PRODUCT AND SPECTRAL PATTERN ANALYSIS CONCEPT - CASE STUDY IN HANOI, VIETNAM

Abstract. Recently USGS released provisional Landsat 8 Surface Reflectance product, which allows conducting land cover mapping over large composed of number of image scenes without necessity of atmospheric correction. In this study, the authors present a new concept for automated classification of land cover. This concept is based on spectral patterns analysis of reflected bands and can be automated using predefined classification rule set constituted of spectral pattern shape, total reflected radiance index (TRRI) and ratios of spectral bands.

Given a pixel vector B₆ = {b₁,b₂,b₃,b₄,b₅,b₆} where b₁, b₂,...,b₆ denote bands 2, 3, ...,7 of OLI sensor respectively. By using the pixel vector B6 we can construct spectral reflectance curve. Each spectral curve is featured by a shape, which can be described in simplified form of an analogue pattern, which is consisted of 15 digits of 0, 1 and 2 showing mutual relative position of spectral vertices. Value of comparison between band i and j is 2 if b_j > b_i, 1 if b_j = b_i and 0 if b_j < b_i. Simplified spectral pattern is defined by 15 digits as m_1,2m_1,3m_1,4m_1,5m_1,6m_2,3m_2,4m_2,5m_2,6m_3,4m_3,5m_3,6m_4,5m_4,6m_5,6 where m_i,j is result of comparison of reflectance between b_i and b_j and has values of 0, 1 and 2. After construction of SSP for each pixel in the input image, the original image will be decomposed to component images, which contain pixels with the same SRCS pattern. The decomposition can be written analytically by equation A = Σⁿ_k=1C_k where A stands for original image with 6 spectral bands, n is number of component images decomposed from A and C_k is component image. For this study, we use Landsat 8 OLI reflectance image LC81270452013352LGN00 and LC81270452015182LGN00. For the decomposition, we use only six reflective bands. Each land cover class is defined by SSP code, threshold values for TRRI and band ratios. Automated classification of land cover was realized with 8 classes: forest, shrub, grass, water, wetland, develop land, barren and others. This paper provides a preliminary research result on application of multispectral image decomposition using simplified spectral pattern for classification of land cover for Hanoi area.