17.3 A Threshold Model for Mapping PML Using MODIS
Time Series Data
17.3.1 Methodology
The decision tree has been proving efficient classifiers for land-cover mapping
(Schneider et al. 2010). However, most decision-tree methods require all classes
that occur in a training image to be exhaustively labeled (Munoz-Marf et al. 2007).
This will not only increase the classification cost since the process of gathering
training samples or otherwise labeling training samples is very expensive in terms of
time and manpower (Byeungwoo and Landgrebe 1999) but also be unnecessary in
those studies in which only a specific class needs to be extracted. Therefore,
one-class classification methods (e.g., Manevitz and Yousef 2001), which try to
detect a specific class and reject the others, have been employed in land-cover
mapping (e.g., Sanchez-Hernandez et al. 2007) and proved to be effective (Foody
et al. 2006; Li et al. 2011).
One example of one-class classification methods is the threshold model, which
can be expressed as:
If (the discriminative features of a pixel meet the threshold conditions) Then
Assign the pixel to the special class;
Else
Assign the pixel to the other class.
To some extent, the threshold model is simple, efficient one-class classifiers
which classify the whole image into the specific class and the other class via
threshold conditions. Therefore, the key to successfully using the threshold model
for land-cover mapping is to correctly select the discriminative features (or simply
the discriminators) and set the correct threshold values for the conditions. In this
study, we applied the threshold model to detect PML from MODIS images since we
just focus on the PML class and ignore the other classes.
Because of its manageable data volume and high temporal resolution (covering
the entire Earth surface every 1–2 days), MODIS images have been widely used in
monitoring regional land surface processes (Stefanov and Netzband 2005; Schaaf
et al. 2002). In this study, both the spectral and temporal features of MODIS images
are explored for determining the discriminators and threshold values for mapping
the PML with the threshold method.
Spectrally, plastic film has very distinct features that can be used to set the
threshold conditions. Figure 1 shows the spectral curves of typical plastics (such
as Clear_PE, Black_PE, and White_PVC, where Clear_PE is similar to TPF) in
comparison with water and vegetation by using USGS Digital Spectral Library
splib06a. From Fig. 17.7, we can find that the TPF has very distinct, high, and
almost constant reflectance from visible to near-infrared wavelengths, which correspond to MODIS band 1 and band 2. Therefore, high reflectance values for these two
bands during the planting stage signify the possible PML with TPF.
17 Remote Sensing–Based Mapping of Plastic-Mulched Land Cover
365
Time Series Data
17.3.1 Methodology
The decision tree has been proving efficient classifiers for land-cover mapping
(Schneider et al. 2010). However, most decision-tree methods require all classes
that occur in a training image to be exhaustively labeled (Munoz-Marf et al. 2007).
This will not only increase the classification cost since the process of gathering
training samples or otherwise labeling training samples is very expensive in terms of
time and manpower (Byeungwoo and Landgrebe 1999) but also be unnecessary in
those studies in which only a specific class needs to be extracted. Therefore,
one-class classification methods (e.g., Manevitz and Yousef 2001), which try to
detect a specific class and reject the others, have been employed in land-cover
mapping (e.g., Sanchez-Hernandez et al. 2007) and proved to be effective (Foody
et al. 2006; Li et al. 2011).
One example of one-class classification methods is the threshold model, which
can be expressed as:
If (the discriminative features of a pixel meet the threshold conditions) Then
Assign the pixel to the special class;
Else
Assign the pixel to the other class.
To some extent, the threshold model is simple, efficient one-class classifiers
which classify the whole image into the specific class and the other class via
threshold conditions. Therefore, the key to successfully using the threshold model
for land-cover mapping is to correctly select the discriminative features (or simply
the discriminators) and set the correct threshold values for the conditions. In this
study, we applied the threshold model to detect PML from MODIS images since we
just focus on the PML class and ignore the other classes.
Because of its manageable data volume and high temporal resolution (covering
the entire Earth surface every 1–2 days), MODIS images have been widely used in
monitoring regional land surface processes (Stefanov and Netzband 2005; Schaaf
et al. 2002). In this study, both the spectral and temporal features of MODIS images
are explored for determining the discriminators and threshold values for mapping
the PML with the threshold method.
Spectrally, plastic film has very distinct features that can be used to set the
threshold conditions. Figure 1 shows the spectral curves of typical plastics (such
as Clear_PE, Black_PE, and White_PVC, where Clear_PE is similar to TPF) in
comparison with water and vegetation by using USGS Digital Spectral Library
splib06a. From Fig. 17.7, we can find that the TPF has very distinct, high, and
almost constant reflectance from visible to near-infrared wavelengths, which correspond to MODIS band 1 and band 2. Therefore, high reflectance values for these two
bands during the planting stage signify the possible PML with TPF.
17 Remote Sensing–Based Mapping of Plastic-Mulched Land Cover
365