control set and no-change set determined from scattergrams (Tokola et al., 1999). Yang
and Lo (2000) found that normalisation methods that used a large number of samples
exhibited a better overall performance, but reduced the dynamic range and coefficient
of variation of the images and therefore reduced the accuracy of image classification.
5.3 CHANGE AND TREND DETECTION TECHNIQUES
The information requirements of the project and the environment of interest guide
the choice of change or trend detection technique. No single change detection technique
is suitable for the myriad of monitoring applications, with the various methods often
giving differing map accuracy (Rogan et al., 2002). Change detection methods include
direct image differencing, spectral index differencing, linear change enhancement
techniques (e.g. selective principal components analysis), direct multi-date
unsupervised classification, post-classification change differencing, and decision tree
analysis (Coppin and Bauer, 1996; Mas, 1999). Typically, these techniques are applied
to imagery collected at two dates, with the differencing and linear change enhancement
techniques resulting in a continuous map product that is subsequently thresholded to
provide change classes. The classification approaches are either applied individually to
each image, where the change can then be classed as change from one cover type one to
another, or to the entire image stack. In this case, the output classification will need
careful interpretation to develop reliable change classes (Jensen, 1996b). Trend
detection methods typically involve the analysis of absolute values of some variable
such as a vegetation index or chlorophyll or TSS concentration, and rely on a form of
per-pixel time-series analysis through fitting of polynomial functions such as Fourier
or wavelet analysis (Ruiz Luna and Berlanga Robles, 1999; Li and Kafatos, 2000;
Coppin, 2003). These deterministic trend detection models are advantageous, since they
can be applied in the same way to a variety of similar trend detection situations,
resulting in standardised reporting of the trend in the indicator of interest in different
regions.
5.4 PRESENTATION OF CHANGE AND TREND DETECTION RESULTS
Accuracy assessment is an important feature of mapping, not only as a guide to
map quality and reliability, but also in understanding thematic uncertainty and its likely
implications to the end user (Czaplewski, 2003). Prior to image classification,
calibration data must be sampled from appropriate areas, at an appropriate support size
(Stehman and Czaplewski, 1998). However, sampling for change detection is more
challenging than that found in single-date approaches (Biging et al., 1998). Typically, a
first step in this process is to highlight areas of change vs. no-change. This can be
accomplished using an optimal threshold value based on similar spectral band
comparisons between dates, vegetation indices, or texture measures (Lunetta et al.,
1998). To ensure appropriate sampling of no-change areas, the stratified adaptive
cluster sampling (SACS) approach has been recommended (Brown and Manly, 1998).
SACS has particular utility for sampling disturbed locations (changed land-cover and
land-use) because they usually represent a minor portion of the target population (most
of the land area has not changed) and are often clustered (Rogan et al., 2002).
Following classification, the accuracy of the change maps must be assessed. The
total error in a thematic map is the sum of the following: 1) reference data errors;
2) sensitivity of the classification scheme to observer variability; 3) inappropriateness
236
Phinn, Joyce, Scarth and Roelfsema
and Lo (2000) found that normalisation methods that used a large number of samples
exhibited a better overall performance, but reduced the dynamic range and coefficient
of variation of the images and therefore reduced the accuracy of image classification.
5.3 CHANGE AND TREND DETECTION TECHNIQUES
The information requirements of the project and the environment of interest guide
the choice of change or trend detection technique. No single change detection technique
is suitable for the myriad of monitoring applications, with the various methods often
giving differing map accuracy (Rogan et al., 2002). Change detection methods include
direct image differencing, spectral index differencing, linear change enhancement
techniques (e.g. selective principal components analysis), direct multi-date
unsupervised classification, post-classification change differencing, and decision tree
analysis (Coppin and Bauer, 1996; Mas, 1999). Typically, these techniques are applied
to imagery collected at two dates, with the differencing and linear change enhancement
techniques resulting in a continuous map product that is subsequently thresholded to
provide change classes. The classification approaches are either applied individually to
each image, where the change can then be classed as change from one cover type one to
another, or to the entire image stack. In this case, the output classification will need
careful interpretation to develop reliable change classes (Jensen, 1996b). Trend
detection methods typically involve the analysis of absolute values of some variable
such as a vegetation index or chlorophyll or TSS concentration, and rely on a form of
per-pixel time-series analysis through fitting of polynomial functions such as Fourier
or wavelet analysis (Ruiz Luna and Berlanga Robles, 1999; Li and Kafatos, 2000;
Coppin, 2003). These deterministic trend detection models are advantageous, since they
can be applied in the same way to a variety of similar trend detection situations,
resulting in standardised reporting of the trend in the indicator of interest in different
regions.
5.4 PRESENTATION OF CHANGE AND TREND DETECTION RESULTS
Accuracy assessment is an important feature of mapping, not only as a guide to
map quality and reliability, but also in understanding thematic uncertainty and its likely
implications to the end user (Czaplewski, 2003). Prior to image classification,
calibration data must be sampled from appropriate areas, at an appropriate support size
(Stehman and Czaplewski, 1998). However, sampling for change detection is more
challenging than that found in single-date approaches (Biging et al., 1998). Typically, a
first step in this process is to highlight areas of change vs. no-change. This can be
accomplished using an optimal threshold value based on similar spectral band
comparisons between dates, vegetation indices, or texture measures (Lunetta et al.,
1998). To ensure appropriate sampling of no-change areas, the stratified adaptive
cluster sampling (SACS) approach has been recommended (Brown and Manly, 1998).
SACS has particular utility for sampling disturbed locations (changed land-cover and
land-use) because they usually represent a minor portion of the target population (most
of the land area has not changed) and are often clustered (Rogan et al., 2002).
Following classification, the accuracy of the change maps must be assessed. The
total error in a thematic map is the sum of the following: 1) reference data errors;
2) sensitivity of the classification scheme to observer variability; 3) inappropriateness
236
Phinn, Joyce, Scarth and Roelfsema