Evaluation
The goal in this section is to quantitatively and qualitatively
present comparative analysis of the proposed procedure with
the other three common unsupervised change-detection meth-
ods from the literature, including two context-insensitive
techniques (namely
PCA
and
FCM
) and one context-sensitive
technique (namely
EM-MRF
).
1. Principle Component Analysis (
PCA
)
Principle Component Analysis is based on transforma-
tion of the multivariate data to several uncorrelated
bands. First, we merged the first three bands of two
Landsat images into six bands and then applied
PC
transform (Deng
et al.
, 2008). The changed information
is usually considered to be in the second component.
Since the histogram distribution of second band pres-
ents a unimodal pattern, two-sided T-point thresholding
is used for separating changed and unchanged region.
2. Fuzzy c-means (
FCM
)
Clustering is one of most common unsupervised tech-
niques for image classification. A powerful technique
from clustering family called fuzzy c-means has been
adopted for unsupervised change detection (Ghosh
et
al.
, 2009). This method is often considered to be more
suitable than hard-membership approach for handling
(a) Landsat image in 1990
(b) Landsat image in 2001
(c) PCA
(d) FCM
(a) EM-MRF
(b) Proposed method
Plate 2. The multi-temporal Landsat TM images [(a) and (b)] and the change detection results from the different unsupervised change-
detection approaches (c to f). The changed pixels are shown in Red. The five yellow dashed rectangles in (a) are the sub-areas that are
illustrated in Figure 4.
642
August 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
