In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
706
piece of Cropland is changed to forest land from one year to
another, we have a chance of 92.6% that the change is true, and
if Cropland is changed to Grassland or Perennial land, the rate
of the mistakes made for detecting the change is also very
small.
Similarly, Forest land (Coniferous, Deciduous, and Mixed
Forest) is almost not confused with Cropland (only 1 out of
1500 forest land samples was classified as cropland). This fact
implies that if Forest land cover is changed to Cropland, we
have a confidence that the change is true. Table 2 also indicates
that Forest land is not mixed much with Perennial and Native
Grassland, with only 1 and 7 out of 1500 were classified as
Grassland, and Perennial land, respectively. Hence, if a piece of
Forest land cover is detected a change to Grassland or
Perennial land, the change is likely true.
Although Grassland and Perennial land are mixed by each
other, they are separable with Forest land, and only 2 and 4 out
of 500 samples of Grassland and Perennial land are classified
as Forest land cover, respectively. This means that if a change
is detected from Grassland or Perennial land to Forest land, the
change is highly likely. The same conclusion can be made for a
change from Grassland or Perennial land to Cropland, although
the confidence is slightly lower as there are 10 out of 500, and
34 out of 500 samples of native Grassland and Perennial land
are classified as Cropland, respectively.
Figure 1 shows the comparison of the reference map and the
classified land cover map using the method discussed above
with the time-series of 10 days cloud-free MODIS composite
data. Figure la is Saskatchewan portion of Circa 2000 land
cover map. The reference map was generated by Agriculture
and Agri-food Canada using 30m Landsat TM data, while
Figure lb is generated by using 250m MODIS data (NDVI +
bands) and the developed method. Although Figure lb is not as
rich in term of spatial details, it is evident that the identified
land cover types are generally agreeable to the Circa 2000 land
cover map.
3.2 Results of homogeneous pixels for training and
dominant pixels for verification
The above assessment and analysis are based on the results
generated using homogeneous samples for training and
verification, which can be applied to the landscape with
homogeneous land cover.
For the evaluation, we used the same homogeneous sample
pixels for model training, but randomly and independently
selected dominant land cover pixels (which could enclose some
homogenous pixels) for verification. Statistically, the
verification samples represent more than 95% situations of the
MODIS pixels of the study region. The same three groups of
variable combinations described above were evaluated for the
process. Table 3 lists the highest accuracy from the three
variable combinations above, respectively.
In comparison with the results of using homogeneous pixels for
training and verification, the three variable combinations using
homogeneous pixels for training and dominant pixels for
verification yield similar but lower accuracy. However, the
overall accuracy of the land cover identification reaches about
-80%. Considering the spatial resolution of MODIS, the results
are encouraging.
Land Cover
Type
Accuracy (%)
Bands
NDVI + bands
Phenology + bands
Cropland
89.20
90.40
88.80
Forest land
87.13
87.73
88.40
Grassland
70.40
70.80
70.80
Shrubland
68.60
70.40
70.00
Perennial
70.00
71.80
73.00
Developed
78.00
77.80
76.40
Water
87.40
89.00
88.60
Average
78.68
79.70
79.43
Table 3. Accurate percent of land cover identification using
dominant pixels for verification
It can be seen from Table 3 and Table 2, Cropland cover type
has similar accuracies from the two methods. This is because
Cropland in Saskatchewan has large parcels. Once Cropland
cover becomes dominant, it is likely that all the 25 sub-pixels
of a MODIS pixel are Cropland cover. Other land covers yield
a lower accuracy (~2%—8% lower) except Perennial land
cover. The results are explainable because the land covers other
than the dominant one within a MODIS pixel would contribute
to the spectral information, and then somewhat confuse the tool
for accurately identifying the dominant land cover. The degree
of the confusion may depend on the number and the types of
land covers within the pixel of the dominant land cover.
An exception is that Perennial land cover which has an equal or
a slightly higher accuracy for NDVI and phenology with bands
combinations. The reason for the higher accuracy needs further
investigation.
4. CONCLUSION AND DISCUSSION
Land cover mapping and its subsequent transitional land
assessment at a regional and national level require large
coverage and adequate spatial and temporal resolutions of EO
data. MODIS data is a reasonable choice. Although promising,
however, based on our evaluations, its usefulness depends on
two critical variables: 1) landscape and 2) spectral
characteristics of targeted objects. Once the targets are
determined, the major factors of affecting land cover
identification are the spatial distribution patterns of land covers.
The degree of landscape heterogeneity under the study area
determines the degree of mixed information within a pixel, and
then plays a major role in affecting the accuracy of land cover
identification.
Our study shows that, in the study region of Saskatchewan,
about 49% of landscape is homogeneous with only one land
cover, and 46% of the land is dominated by one land cover
based on the size of a MODIS pixel and Circa 2000 land cover
map. The accuracies of the land cover identification for the
homogeneous and dominant landscape (includes homogeneous
landscape) are about 88% and 80%, respectively. These suggest
that MODIS may provide valuable information for the
transitional land mapping for Saskatchewan region although
further evaluation is needed such as improving the quality of
the time-series of MODIS data and the method developed. For
