17.3.2.2 Determination of Threshold Condition and Value
There are multiple temporal-spectral features of MODIS time series usable as the
discriminator for PML mapping. In order to determine the best one as the discriminator as well as its associated d and x, we need to analyze the performance of
different features. By using PML and non-PML training sets, we calculated the
means of band 1 reflectance, band 2 reflectance, and NDVI for PML and non-PML
for each day. Therefore, three pairs of daily-mean-spectral-reflectance (DMSR)
curves, which are band 1, band 2, and NDVI, from MODIS time series for each year.
Figure 17.8 shows the results of the comparison of the DMSR time series between
PML and non-PML using original MODIS data in 2009, 2013, and 2014. From
Fig. 17.8, we can find that in all 3 years, band 1 reflectance of PML is generally
higher than non-PML while both band 2 and NDVI are just opposite. However, there
were 35 days in 2009, 43 days in 2013, and 35 days in 2014 when values of band
1, band 2, and NDVI for both PML and non-PML are very close. By checking the
quality control (QC) layer of MODIS data, it was found that the dates, when the
values for both PML and non-PML were very close, were covered by cloud either
fully or partly.
In order to remove the impact of cloud in this study, we linearly interpolated the
pixel value for the day when the MODIS QC layer indicates the pixel was covered
either fully or partly by cloud from the nearest earlier and later good days. After the
interpolation, the means were recalculated to reconstruct DMSR for both Band 1 and
NDVI time series (Fig. 17.9). Band 2 time series was not reconstructed because there
was no significant difference in band 2 between PML and non-PML. The reason for
no significant difference is that the non-PML area in this study was mainly winter
wheat which has a high NIR reflectance in May.
Figure 17.9 show that, from the early sowing season to the end of the early
growing season (the detectable period), the DMSR of band 1 for PML is obviously
greater than that of non-PML (winter wheat and spring cornfields). The reason is
that, from farmers who put the new transparent plastic mulch on the fields in the
sowing season, the plastic mulches, which has higher spectral reflectance in band
1, are visible in the early growing season. By further analyzing Fig. 17.9a, b, c, we
find that PML DMSR curves of band 1 fluctuate around 0.2 while non-PML 0.15 and
this pattern of difference were the same for all 3 years. Figure 17.9d, e, f illustrate
that NDVI values of non-PML are above that of PML in 3 years, since cotton
phenology differs from that of winter wheat, namely, winter wheat is in the greenup stage while cotton in sowing and emerging stage. Figure 17.9d, e, f also show that
NDVI values for non-PML in the 3 years of study fluctuated between 0.2 and 0.5 but
generally went up with the progress of the growing season. However, the NDVI
values for PML can be divided into two stages. The first stage is from the 90th day to
the 125th day when NDVI fluctuated between 0.1 and 0.25 and the second stage
from the 126th day to the 150th day when NDVI went up with the progress of the
season. Therefore, it is clear that we can use NDVI values from the 90th to 125
th day
(total of 31 days), the sowing and seedling emergence periods of cotton, to discriminate PML and non-PML.
368
L. Lu
There are multiple temporal-spectral features of MODIS time series usable as the
discriminator for PML mapping. In order to determine the best one as the discriminator as well as its associated d and x, we need to analyze the performance of
different features. By using PML and non-PML training sets, we calculated the
means of band 1 reflectance, band 2 reflectance, and NDVI for PML and non-PML
for each day. Therefore, three pairs of daily-mean-spectral-reflectance (DMSR)
curves, which are band 1, band 2, and NDVI, from MODIS time series for each year.
Figure 17.8 shows the results of the comparison of the DMSR time series between
PML and non-PML using original MODIS data in 2009, 2013, and 2014. From
Fig. 17.8, we can find that in all 3 years, band 1 reflectance of PML is generally
higher than non-PML while both band 2 and NDVI are just opposite. However, there
were 35 days in 2009, 43 days in 2013, and 35 days in 2014 when values of band
1, band 2, and NDVI for both PML and non-PML are very close. By checking the
quality control (QC) layer of MODIS data, it was found that the dates, when the
values for both PML and non-PML were very close, were covered by cloud either
fully or partly.
In order to remove the impact of cloud in this study, we linearly interpolated the
pixel value for the day when the MODIS QC layer indicates the pixel was covered
either fully or partly by cloud from the nearest earlier and later good days. After the
interpolation, the means were recalculated to reconstruct DMSR for both Band 1 and
NDVI time series (Fig. 17.9). Band 2 time series was not reconstructed because there
was no significant difference in band 2 between PML and non-PML. The reason for
no significant difference is that the non-PML area in this study was mainly winter
wheat which has a high NIR reflectance in May.
Figure 17.9 show that, from the early sowing season to the end of the early
growing season (the detectable period), the DMSR of band 1 for PML is obviously
greater than that of non-PML (winter wheat and spring cornfields). The reason is
that, from farmers who put the new transparent plastic mulch on the fields in the
sowing season, the plastic mulches, which has higher spectral reflectance in band
1, are visible in the early growing season. By further analyzing Fig. 17.9a, b, c, we
find that PML DMSR curves of band 1 fluctuate around 0.2 while non-PML 0.15 and
this pattern of difference were the same for all 3 years. Figure 17.9d, e, f illustrate
that NDVI values of non-PML are above that of PML in 3 years, since cotton
phenology differs from that of winter wheat, namely, winter wheat is in the greenup stage while cotton in sowing and emerging stage. Figure 17.9d, e, f also show that
NDVI values for non-PML in the 3 years of study fluctuated between 0.2 and 0.5 but
generally went up with the progress of the growing season. However, the NDVI
values for PML can be divided into two stages. The first stage is from the 90th day to
the 125th day when NDVI fluctuated between 0.1 and 0.25 and the second stage
from the 126th day to the 150th day when NDVI went up with the progress of the
season. Therefore, it is clear that we can use NDVI values from the 90th to 125
th day
(total of 31 days), the sowing and seedling emergence periods of cotton, to discriminate PML and non-PML.
368
L. Lu