ACCURACY EVALUATION OF DIGITAL SURFACE MODEL USING ALOS PRISM
DATA FOR DISASTER MONITORING
Kyaw Sann 00, Takahiro NAKAGAWA and Masataka TAKAGI
Department of Infrastructure Systems Engineering, Kochi University of Technology
Tosayamada, Kochi 782-8502. Japan
Tel (81)-887-53-1040 Fax: (81)-887-57-2420 -
108007u@gs.kochi-tech.ac.jp, 115114y@gs.kochi-tech.ac.jp, takagi.masataka@kochi-tech.ac.jp
Commission IV, WG-IV-9
KEY WORDS: Remote Sensing, DSM, Image Matching, Accuracy Measurement, Georeferencing, Disaster Monitoring
ABSTRACT:
The need of accurate measurement is a tactical objective to produce precise information in disaster monitoring. A small movement of
land could be a prior index to the biggest landslide. The biggest disasters can lead by a small unnoticeable changes. The
measurement of 3 dimensional surfaces changes could be calculated accurately using very high resolution satellite stereo images.
The availability of short-time periodic satellite’s stereo-pair images are benefits to investigate the land movements when the images
had acceptable accuracy. Vendors are providing camera replacement models with a set of rational polynomial coefficients (RPCs) to
replace physical camera models for registration improvement. Fraser & Hanley (2003) represented that no loss in accuracy is to be
expected when bias corrected RPCs are used for georeferencing. Even if RPC coefficients are provided by vendors, those are not
enough to investigate small changes of land-form, slope displacements. Thus, we should be applied a pliant sensor model based on
very accurate ground control points (GCP). First of all, high resolution digital surface model (DSM) is generated from ALOS PRISM
triplet images; resulting X=0.97, Y=0.99 and Z=1.85 pixels in accuracy. This result was produced by stereo matching between nadir
and backward looks. Finally, the accuracy evaluation of 3D measurement is conducted based on selected check points (CP). Afterall,
error vectors will be produced by validating with Airborne Laser Scanner DSM data. Even though the result measurement will be
biased with some errors, the future work could be improved by matching a series of triplet images.
1. INTRODUCTION
In disaster monitoring, accurate 3 dimensional surfaces
measurement is a tactical objective to notice land movements
and land-form changes. A small movement of land could be
accumulated to form the biggest movement. Thus, the biggest
disaster can be lead by a small unnoticeable land-forms change.
The detection of the changes could be calculated accurately,
when we used accurate input to the analysis process. Using
very high resolution satellite images we could be detected small
changes, if our input has potential in ground accuracy.
Nowadays, there are a lot of very high resolution satellite
images are available by the advancement of space satellites
mounting optical sensors. Vendors are providing camera
replacement models with a set of rational polynomial
coefficients (RPCs) to replace physical camera models for
registration improvement. Fraser & Hanley (2003) represented
that no loss in accuracy is to be expected when bias corrected
RPCs are used for georeferencing. Even if RPCs and alternative
affine projected model are provided by vendors, some errors are
still contaminating in the images. The errors could not cover for
the sensor alignment. Ground control points (GCP) are
necessary in geometric correction to precise (Hashimoto, 2006).
In this study, we used GCP data acquired by GPS-VRS. It is a
acronym of differential GPS based on Virtual Reference
Stations.
First of all, it is need to generate high resolution digital surface
model (DSM) from available satellite imagery. Using ALOS
PRISM triplet data it could be possible to generate high
resolution DSM data. In this study, we used triple image of
ALSO PRISM to extract DSM. The image had such three looks
as forward, nadir and backward; this conjugated triplet image
could be input to generate DSM when the images are matched.
The prior study found that ALOS PRISM data is contaminated
with some systematic errors on along track of sensors. Those
errors are accumulated in the forward and backward looks. The
georeferencing result from vendors provided RPC model is 1.3
x 1.8 pixels accuracy in nadir, 7.11 x 1.8 for forward look and
8.4 x 0.9 for backward look. Based on result values, we found
that more errors are in forward and backward and in along track
direction. In this study, we corrected this contaminated
systematic errors by applying 3D perspective projection.
Elevated mountains, river low lands and farm plains are
distributed on the selected study area. The situation of the study
area is supporting to use 3D perspective projection for
georeferencing. After applying the perspective projection
model, the calculation produced accuracy improvement. Result
values are in nadir 0.26 x 0.29 pixels, in forward look
0.26x0.27 pixels and in backward look 0.21 x 0.25 pixels
respectively. The result of 3D perspective projection give better
accuracy than vendor provided RPC models. Even if the
accuracy of result is improved, the study agreed to provide very
high accurate result. For the purpose, a corrected RPC model
for triplet images was generated using ground control points
(GCP). GCP are measured by GPS VRS observation to get very
precise measurement. GPS VRS is a latest technology of
location measurement; it is based on differential GPS with
virtual reference station using wireless telecommunication. This
measurement produced very high accurate location in collection
of GCP. Then, stereo matching was conducted by least squares
matching using available low resolution digital elevation model
