Short- Term change detection with precise geometric correction and sub-pixel land cover characterization of modis
Methodology
1. Examination of geometric correction accuracy
Detecting change in sub-pixel level requires precise overlay between different images. It depends on
the accuracy of geometric correction of each image. Firstly, the accuracy of geometric correction done by
WebMODIS is examined by pattern matching. NDVI images were used to conduct pattern matching. As
matching piece should have a characteristic shape, Oshima and lake Kasumigaura in the study area are
selected. NDVI images of these two areas are shown in Fig.3. Correlation coefficients are calculated
between the matching piece and 300 x 300 km
2 coverage.
Fig.3. NSM images of matching pice (left:Oshima, right:Kasumigaura)
Fig.4. Process of pattern matching
2. Linear mixture modeling
MODIS is one of the few space-borne sensors currently capable of acquiring radiometric data over a
broad range of view angles. However, the relatively coarse spatial resolution of the MODIS most often
results in measurements of mixed land covers, and thus the pixel unmixing is indispensable. A linear
spectral mixture model based on three end-members including vegetation, soil and water is defined by the
equations and constraints below
where MODIS
ch1 , MODIS
ch2 are the MODIS radiance of visible and near infrared channels respectively,
and V, S and W are the fractional coverage of vegetation, soil and water respectively. ij a are the
end-members and range from 0 to 1.
The number of equations equal to three by incorporating the constraint equation (Eqn.3) in order to
calculate the fractional coverage V, S and W. Eqn.3 represents that the sum of three fractions in one pixel
equals to 1. By solving three equations (Eqn.1, 2 and 3) simultaneously, we can calculate the fractional
coverage of V, S and W.
3. Endmember selection
A mixture model based on three end-members has the simple geometrical interpretation as the triangle
whose vertices are the end -members (Fig.5). Since a finite number of hull facets has only finitely many
partitions, there are only finitely many simplexes the method can construct, and finding the one with the
smallest volume is a combinatorial optimization problem. The triangle is defined to include 95 % of convex
hull to eliminate the exceptional noisy pixels. Fig.6 shows the scatter plot for red and near infrared spectral
distribution of MODIS image acquired on 10 Mar. 2002. The vertices of the triangle is defined as follows
- Vegetation : highest value in near infrared reflectance
- Soil : highest value in red reflectance
- Water : lowest value in both red and near infrared reflectance.
Fig 5. Relationship between VSW indices end-members triangle for Red-NIR
Fig 6. Scatter plot for Red-NIR on 2002.3.10
4. Change detection model and evaluation of land- cover change
Fractional coverage of vegetation, soil and water will be derived by spectral mixture analysis. In this
study, we assume that there is no land cover change between the two scenes. In addition to that, it is
assumed that there is mis-registration error of only one pixel between the two images. Using this
assumption, a 3 x 3 filter as described hereafter is applied to reduce the errors. The spatial filtering
will be applied to the V, S, and W images on 12 Mar. 2002. Fig.7 depicts the algorithm of the spatial filtering.
The process is as follows:
- Select pixel in Image2 in around 9 pixel of (i, j) involved itself, such as minimize difference between
the value and other pixel value (i , j) in Image1,
- The value of pixel (i , j) in Image2 is replaced selected pixel value,
- All pixel was done by these transaction and get Image2’ .
Then, three filtered images of vegetation, soil and water on 12 Mar. 2002 will be derived from this
processing. Land cover change will be evaluated by simple subtraction between the 10 Mar. 2002 and the
12 Mar. 2002 images of each three end -members. The pixel value of subtracted vegetation, soil, and water
image measures land cover change.
Fig 7. Filtering algorithm
Results and Discussion
1. Accuracy assessment of geometric correction by WebMODIS
Correlation coefficients were calculated between the matching pieces and the whole images of both
MODIS data on 10 Mar. 2002 and 12 Mar. 2002. Fig.8. shows the result of the pattern matching. Best
matching points are easily distinguished from the nearby pixels in these four pairs. Best matching points for
both Kasumigaura and Oshima were same. This result shows that the accuracy of geometric correction by
WebMODIS is high in pixel level.
Fig 8. Result of pattern matching
Table1. RMSE of percentage for filtered and unfiltered image
