Research Methodology - Historical Land Use/Land Cover Classification Using Remote Sensing - page 77

MAD can be used for bi-temporal change detection and for automatic relative
radiometric normalization (Nielsen 2007 ; Canty and Nielsen 2008 ; Canty 2010 ).
Canty ( 2010 ) explained the mathematical background of MADs.
To create a MAD-image, it is necessary to select two multi-spectral images that
have alike spatial dimensions (size of the pixels). The two images will be modeled
as a casual variable G1 and G2. When each image has, for example, 123 pixels,
then these 123 pixels have a 123 times repetition of a mathematical random
experiment, where,—here, the accurate value of pixels are not defined or descri-
bed. If G1, G2 represent only a specific pixel or an entire image, then it will be
illogical for them. What is important here is the properties of the causal variables
G1, G2. Some suppositions about G1, G2 can be made by using the metrics of
histogram (e.g., empirical variance, mean-based assessment of predictable value).
Each image includes an N spectral bands, with G1 (also G2) as a random vector
(Schultz 2011 ).
The X 2 image expresses the representative pixels which may be suitable for the
radiometric normalization (Canty 2009 ). The X 2 distributions are only the pixels
that satisfy the formula: /Pr (no change) [ t, where (t) is a decision threshold that
is typically 95 %/. The radiometric normalization based on these satisfying pixels
will be used to perform an orthogonal regression.
The iMADs, X 2 -values can only be calculated for overlapping areas, since the
iMAD is designed for applying the automated radiometric normalization of multi-
temporal remotely sensed data sets.
Adjacent scenes can be normalized by selecting their overlapping area (subsets)
and followed by using the created transfer function of the orthogonal regression
expressed on an entire image. It is important to cover all LULC-properties in the
overlapping region of the two images (master and target), while pixels with an
alike spectral behavior from overlapping and non-overlapping regions will be
treated according to the regression function (Canty and Nielsen 2008 ).
Large water bodies affect the iMAD negatively (Canty 2009 ). Clouds and their
shadows do not affect the normalization superiority, while they are detected as
change (Canty and Nielsen 2008 ).
Summarized after Canty and Nielsen ( 2008 ) and Schultz ( 2011 ), the performed
radiometric normalization was achieved in the five phases: (1) insert the dual-
temporal data set; (2) compute CVs, build MADs and reweighing the spectral
information accordingly; (3) repeating until no significantly improvement in cor-
respondence of the CVs; (4) select pixels that have a no-change chance greater
than a threshold value (t); and (5) determine the two radiometric normalization
coefficients, i.e. slope and intercept, based on the orthogonal regression on selected
pixels that have to be performed previously. The iMAD was applied to the imagery
using ENVI 4.6 and IDL 7.06. The source code used was provided by Morton
Canty and can be downloaded at '' http://mcanty.homepage.t-online.de/
software.html ''. In Canty ( 2009 ) the implementation and installation of the soft-
ware to ENVI 4.6 is presented and explained.
To normalize the radiometry of all the used remote sensing sensor (e.g.,
LANDSAT-MSS-June-1975

and LANDSAT-TM-August-2007) data sets, a