In: Wagner W„ Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
developed, and the results produced, as well as give a brief 
discussion. 
2. DATA AND METHOD 
MODIS sensors aboard the Terra and Aqua satellites have a 
revisiting frequency of 1 to 2 days, acquiring data in 7 spectral 
bands for land surface applications, with a spatial resolution 
ranging from 250 - 500 metres. It is one of the most advanced 
sensors available for large-scale terrestrial applications 
(Salomonson et al., 1989). To facilitate use of the datasets, 
Canada Centre fore Remote Sensing has developed a new 
technology to product of 10 days cloud-free composites of 
MODIS 7 land bands covering Canada and North America. 
Additional to other characteristics of this product, bands 3-7 are 
downscaled from the original spatial resolution of 500 meters 
to 250 meters (Luo et al., 2008). Therefore all the 7 bands of 
the cloud-free composites have the same spatial resolution and 
dimension. 
For the study, only the data in the vegetation growing season 
(April to October) of the time series of MODIS 10 days cloud- 
free composites are analyzed. For the selected datasets there are 
3 composites every month, and in total 21 composites are 
available from the beginning of the vegetation growing season 
to the end of the growing season. Considering all the 7 bands 
for each composite, there is a total of 147 ‘bands’ at various 
periods of the vegetation growing season. 
Two additional datasets, NDVI and phenology in the 
vegetation-growing season, are derived from the time series of 
MODIS bands. NDVI, representing ‘greenness’ of vegetation, 
has been widely used for vegetation mapping. In total, same as 
the number of the MODIS bands, there are also a time series of 
147 NDVI values in the vegetation growing season as they are 
derived from MODIS band 1 and band 2. 
Another derived dataset from the MODIS data is vegetation 
phenology parameters. Phenology represents vegetation 
periodic biological phenomenon, which can be derived from 
NDVI time series by function fitting. There are total 11 
vegetation phenology parameters including starting and ending 
of growing season, seasonal amplitude, seasonally integrated 
NDVI, rate of increase at the beginning of the season, rate of 
decrease at the end of the season, etc. Vegetation phenology is 
derived from the MODIS data by the aid of TIMESAT 
(Jonsson and Eklundh, 2003) software. 
The rationale of using NDVI and vegetation phenology to aid 
land cover information extraction is that different vegetation 
types at different growing stages may have distinctive NDVI 
and phenology phenomenon. 
With all the information available (MODIS 10 days cloud-free 
composite bands, NDVI, and phenology parameters), all 
possible data combinations for maximizing the overall land 
cover identification accuracy for targeted land cover types are 
sought and assessed by using See5/C5.0 data mining tool 
which has been used or discussed by various studies (Keane et 
al., 2004; Pal and Mather 2003). 
Year 2000 is the first year when MODIS was operational; also 
there is a good reference of Circa 2000 land cover map (Figure 
la shows the reference land cover map of Saskatchewan, the 
case study province of the activity) which was generated from 
Landsat TM images. Therefore year 2000 is set as the starting 
point for transitional land assessment. 
10 major land cover types of Circa 2000 land cover map 
(Agriculture and Agri-Food Canada, 2008) were mapped in the 
study. They are Annual Cropland, Water bodies, Developed 
land, Native Grassland, Shrubland, Perennial (crop and 
pasture), Wetland, Deciduous, Coniferous and Mixed Forest. 
From bioenergy land cover mapping point of view, it is not 
critical to separate Deciduous, Coniferous, and Mixed Forest 
land cover. They are grouped into Forest land cover type in the 
study. Wetland is conservative land, so it was masked out 
before the mapping. All the analysis and results described 
below are based on the redefined types for the land cover 
mapping. 
It is found that point samples like ground reference collected 
from field survey using GPS are not the best representations for 
model training and for verification of results derived from 
MODIS images due to the relative coarse spatial resolution of 
the data. Instead, area sample method was used in the study. 
Area sampling means that ground reference is established not 
by a point data, but by an area which has a homogenous land 
cover. 
A homogeneous pixel (area) means that the ground represented 
by the pixel on the image is covered by only one land type. 
Truly, land cover is not always homogeneous, and 
heterogeneity is universal. However, homogeneity and 
heterogeneity are relative terms. When they are applied to EO- 
based applications, they are determined by spatial resolution of 
images in relation to the size of features on the ground. A piece 
of land is heterogeneous for a coarse resolution image, but may 
be homogeneous for all its sub-areas with a finer spatial 
resolution imagery (also depends on object size). Homogeneity 
and heterogeneity are also affected by the level of a land cover 
classification system. 
For the study, three types of landscape settings are considered 
within the dimension of a MODIS pixel: they are homogenous, 
dominant and heterogeneous. As described above, a 
homogeneous pixel implies that the ground represented by the 
pixel on the image is covered by only one land type; a 
dominant pixel means that over a half of the ground area is 
occupied by one land cover with other land covers mixed; and 
the rest are heterogeneous cases. 
To identify and evaluate the distribution of these types within 
the study area, the reference Circa 2000 land cover map is 
geometrically matched to and superimposed on the MODIS 
images. A MODIS pixel then corresponds to 25 pixels of circa 
2000 land cover map (which was rescaled from 30 m to 50m). 
A MODIS pixel is classified as homogeneous if and only if all 
the corresponding 25 sub-pixels on the Circa land cover map 
have the same land cover, or dominant if a land cover type (the 
dominant one) has more than 13 or more sub-pixels, or 
heterogeneous (other land cover combinations). 
704