Geography Reference
In-Depth Information
Many unsupervised change detection approaches deal with the multispectral
images to produce an additional image. The most essential basis for these algorithms
is the determining of the finest global threshold in the histogram of the so-called
generated difference image, where the classifying of change and unchange classes is
made on the importance of the resulting spectral change vectors by applying of
empirical or theoretical well-founded global threshold strategies. The best global
threshold depends on the statistical irregularity of the two images, which are often
unknown. Pacifici et al. (
2007
) reviewed the published techniques in the past decade:
the Image Differencing (ID), Normalized Difference Vegetation Index (NDVI),
Change Vector Analysis (CVA), Principal Component Analysis (PCA), Image
Rationing (IR), Expectation Maximization (EM) (Bruzzone and Fernàndez-Prieto
2000
), Markov Random Field (MRF) (Bruzzone and Fernàndez-Prieto
2000
),
Object-Level Change Detection (OLCD) (Hazel
2001
), Reduced Parzen Estimation
(RPE) (Bruzzone and Fernàndez-Prieto
2002
), Maximum a Posteriori Probability
(MPP) decision criterion (Kasetkasem and Varshney
2002
), Multivariate Alteration
Detection (MAD also called Iteratively Reweighted MAD (IR-MAD)) (Nielsen
2007
), MAD and the combined MAF/MAD (Maximum Autocorrelation Factor)
transformations, and Genetic Algorithm (GA) (Celik
2010
).
The above techniques generally do not aim to identify clearly what types of
LULC-changes have taken place in an area (e.g., which vegetated areas have been
urbanized). They are suitable for applications such as detection of burned areas, or
detection of deforestation. However, they are not useful when it is necessary to
define the types of changes that have occurred in the studied area, for example, in:
observing the shifting in cultivation; urban growth; or where it is required to know
all the types of changes that occurred in investigated area.
Advantages: (1) pre-classification is not necessary, so, avoiding the tiring in
classification process at the starting; (2) it is regarded as simple and rapid, and can
be applied on a great number of images; and (3) the ease in fine-tuning to detect
the specific interested changes, and they are, in general, likely to have a higher
ability to find slight changes (Yuan et al.
2005
). Disadvantages: (1) the detection of
image changes, especially if focused on agricultural areas, may be affected by
troubles with phenology and cropping. Such troubles could be worsened by
inadequate image accessibility and poor quality in moderate zones, and the
problems in adjusting poor images (Blaschke
2005
); (2) also, these techniques are
corrupted by: changes in illumination at two times, changes in atmospheric con-
ditions, and in technical sensor calibration. These make complex a direct evalu-
ation between raw imagery obtained at different times where additional processing
steps are required (e.g., radiometric calibration) (Pacifici et al.
2007
); and (3) there
remains the problem of defining the threshold value at which the change between
the two images is measured. Also, it is clear that using unsupervised methods is
obligatory in many remote-sensing applications, when appropriate ground truth
information is not always available (Bruzzone and Fernàndez-Prieto
2002
).
2. Techniques which first assign classes and then detect change (post-classifica-
tion comparison) Supervised Approach Post classification methods.
Search WWH ::
Custom Search