To determine the best d value, there were 31 training sessions with d from 1 to 31.
For each training session, the accuracy of PML detection was calculated by using the
testing sets of ground truths. Then d value with the highest detection accuracy was
selected as the d value of the model.
To sum up, the PML curves in NDVI of 3 years are all under non-PML, despite
their different ranges of DMSR time series, and there is a significant difference
(approximately 0.1) between PML and non-PML of DMSR from 90th to 125th in
3 years; therefore, NDVI DMSR from 90th to 125th is the best discriminator to
detect PML among the three time series.
17.3.2.3 Detecting and Mapping PML
The statement that NDVI DMSR curves for PML fluctuate around 0.2 in 3 years
doesn’t mean that NDVI threshold value of detecting PML should be set to “x < 0.2.”
The reason is that DMSR values are the average spectral reflectance values of all
pixels in the training sets, which means some PML pixels may have NDVI values
greater than 0.2. In other words, this threshold value will lead to misclassification of
PML pixels.
Therefore, we used a threshold model to identify non-PML in the study area and
deduct non-PML from croplands land-cover image to get PML areas. According to
the above analysis and referring to Fig. 17.9d, e, f, among the threshold model, x can
be 0.2. Therefore, the threshold model for PML detection can be written as:
If (the number of the accumulated days > d when the NDVI value of a pixel in
MODIS time series
< 0.2 during 95
th to 125
th day of a year) Then
Assign the pixel to non-PML class;
Else
Assign the pixel to PML class
We developed a java program of the threshold model to detect PML from MODIS
NDVI time series from 90th to 125th. Using PML and non-PML ground truth, which
were derived from Landsat-8 OLI imagery with maximum likelihood classifier and
manual calibration as ground truths, the classification results were evaluated by
overall accuracy (OA) and Kappa coefficient, κ (Cohen 1960) in ENVI software,
and the results of the threshold model Experiments were given in Table 17.4.
Table 17.4 shows that, in 2009, when threshold parameter x is 0.2 and the
accumulated days parameter d is 7, the classification with the threshold model
method has achieved the highest accuracy with OA 0.848 and κ0.660, while in
2013 and 2014 when threshold parameter x is 0.2 and the accumulated days
parameter d is 9 and 8, respectively, classification accuracy reaches their peak with
OA 0.898 and 0864 and κ 0.796 and 0.679, respectively. It also shows that we can
use the same threshold model to detect PML with high accuracy in three diverse
years with x ¼ 0.2 and d ¼ 8. OA values in all 3 years are larger than 0.84 and κ large
17 Remote Sensing–Based Mapping of Plastic-Mulched Land Cover
371
For each training session, the accuracy of PML detection was calculated by using the
testing sets of ground truths. Then d value with the highest detection accuracy was
selected as the d value of the model.
To sum up, the PML curves in NDVI of 3 years are all under non-PML, despite
their different ranges of DMSR time series, and there is a significant difference
(approximately 0.1) between PML and non-PML of DMSR from 90th to 125th in
3 years; therefore, NDVI DMSR from 90th to 125th is the best discriminator to
detect PML among the three time series.
17.3.2.3 Detecting and Mapping PML
The statement that NDVI DMSR curves for PML fluctuate around 0.2 in 3 years
doesn’t mean that NDVI threshold value of detecting PML should be set to “x < 0.2.”
The reason is that DMSR values are the average spectral reflectance values of all
pixels in the training sets, which means some PML pixels may have NDVI values
greater than 0.2. In other words, this threshold value will lead to misclassification of
PML pixels.
Therefore, we used a threshold model to identify non-PML in the study area and
deduct non-PML from croplands land-cover image to get PML areas. According to
the above analysis and referring to Fig. 17.9d, e, f, among the threshold model, x can
be 0.2. Therefore, the threshold model for PML detection can be written as:
If (the number of the accumulated days > d when the NDVI value of a pixel in
MODIS time series
< 0.2 during 95
th to 125
th day of a year) Then
Assign the pixel to non-PML class;
Else
Assign the pixel to PML class
We developed a java program of the threshold model to detect PML from MODIS
NDVI time series from 90th to 125th. Using PML and non-PML ground truth, which
were derived from Landsat-8 OLI imagery with maximum likelihood classifier and
manual calibration as ground truths, the classification results were evaluated by
overall accuracy (OA) and Kappa coefficient, κ (Cohen 1960) in ENVI software,
and the results of the threshold model Experiments were given in Table 17.4.
Table 17.4 shows that, in 2009, when threshold parameter x is 0.2 and the
accumulated days parameter d is 7, the classification with the threshold model
method has achieved the highest accuracy with OA 0.848 and κ0.660, while in
2013 and 2014 when threshold parameter x is 0.2 and the accumulated days
parameter d is 9 and 8, respectively, classification accuracy reaches their peak with
OA 0.898 and 0864 and κ 0.796 and 0.679, respectively. It also shows that we can
use the same threshold model to detect PML with high accuracy in three diverse
years with x ¼ 0.2 and d ¼ 8. OA values in all 3 years are larger than 0.84 and κ large
17 Remote Sensing–Based Mapping of Plastic-Mulched Land Cover
371