x, 99.3%
S
5 99.0%
=
2
= 97 À
€ 0 - SR (8000 clusters)
2 —e— NBAR (8000 clusters)
95 T T T 1
2 3 4 5
Time Separation Delta-t [months]
Figure 9. Classification accuracies obtained by
time-domain co-occurrence matrix which is
defined with 8000 spectral clusters.
The time-domain co-occurrence matrix defined with 8000
spectral cluster shows 396-496 higher classification accuracy
than that of the time-domain co-occurrence matrix defined with
reflectance.
MODI2QI and MCDI12Q1 of MODIS land cover product
which are respectively produced from SR and NBAR products
showed classification accuracy of 81% and 96%, respectively.
That is, the classification accuracy obtained by the time-domain
co-occurrence matrix which is defined with spectral cluster is
3%-18% higher than that of MODIS land cover products.
Classification accuracies of MODIS land cover product were
measured by using same test samples. Because test samples
were extracted from training area for classification of SR and
NBAR products, it is fundamentally presumed that the accuracy
of MOD12Q1 and MCD12Q1 products is lower than that of our
classification results. However, we consider that these
classification accuracies of SR and NBAR products showed
good performance of the proposed simple classification method.
Figure 10 shows the classification results obtained by cosine
distance classifier in the case of the highest accuracy.
5. CONCLUSIONS
Land cover classification for global scale were performed by
using two kinds of multi-temporal MODIS reflectance products.
The proposed method using the time-domain co-occurrence
matrix and the non-parametric minimum distance classifier
showed good classification performance compared with
MOD12Q1 and MCD12Q1 MODIS land cover product.
The highest classification accuracy was obtained when the non-
parametric cosine distance classifier was driven by the time-
domain co-occurrence matrix defined with spectral cluster and
three or four months time-separation. And also, it was cleared
that Surface Reflectance 8-Day L3 product and Nadir BRDF-
Adjusted Reflectance product showed similar classification
accuracy of 99% for IGBP-17 land cover categories.
Future study should be carry out in our classification scheme in
order to examine stability of classification for multiple years
and to validate classification accuracy with more suitable test
samples.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B8, 2012
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia
(a) SR product
(b) NBAR product
Figure 10. Land cover classification results.
Acknowledgements
This study was supported by the JAXA GCOM-C project under
contract "JX-PSPC-332027".
REFERENCES
Alan Srahler, Doug Muchoney, etc., 1 May 1999. MODIS Land
Cover Product Algorithm Theoretical Basis Document (Version
5.0)
E. F. Vermote and S. Y. Kotchenova, March, 2008. MOD09
(Surface Reflectance) Users Guide. Land Processes
DAAC(EROS) in collaboration with the Department of
Mathematics and Computer Science, South Dakota School of
Mines and Technology, January 2006. MODIS Reprojection
Tool User's Manual(Release 3.3a).
Mikio Maeda, Kiyonari Fukue and Haruhisa Shimoda, 2010.
Continental scale land cover classification using MODIS
surface reflectance product. In: Proc. The 31th Asian
Conference on Remote Sensing, Hanoi, Vietnam, pp.PS01-10-1-
PS01-10-6..
Haruhisa — Shimoda and Kiyonari Fukue, 2011.
CONTINENTAL SCALE LAND COVER CLASSIFICATION
USING MODIS SURFACE REFLECTANCE PRODUCTS. In:
2011 IEEE International Geoscience & Remote Sensing
Symposiums, Vancouver, Canada, pp. 692-695.
Haruhisa Shimoda and Kiyonari Fukue, 2011. GLOBAL LAND
COVER CLASSIFICATION USING MODIS SURFACE
REFLECTANCE PRODUCTS. In: Proc. The 32nd Asian
Conference on Remote Sensing, TAIPEI, pp.P_241_8-23-20-1-5.
Inte
SE
KEY
ABS]
Soil e
other
northe
outlet
From
let of
basin
soil c
manag
(78.0¢
found
terrair
Land
water
the nm
additi
soil «
depos
effect
destru
Irriga
geogr
degra
the ef
to 7
loss h
Europ
and S
1993)
In ass
proble
area (
adopt
condit
are ar
they a
2000)
area s
erosio
global
existit
signif