SCS+C TOPOGRAPHIC CORRECTION TO ENHANCE SVM CLASSIFICATION ACCURACY

1
SCSC TOPOGRAPHIC CORRECTION TO ENHANCE SVM
CLASSIFICATION ACCURACY Jwan Al-doski, Shattri
B. Mansor, H'ng Paik San and Zailani Khuzaimah
5-6/11/2019
2
Introduction
When the research region is positioned in rough
or mountainous areas, the significant part of
satellite data pre-processing is a topographic
correction 1. There are many topographical
correction methods (empirical and non-empirical
parametric models). They have been extensively
used and shown that adjustment of topographical
consequences performed prior to the use of
multispectral and multi-temporal pixel
classification can significantly enhance
classification accuracy, especially if the
location of the research is in rough terrain 3.
Topographic correction models are diverse and the
choice of one model relies on its effectiveness
to decrease the relief impact and ease of
application and on the performance and study
region of remote sensing data. Although, it is
pointless to compare the studies since the input
files and parameters are different and depend on
in study areas, vegetation types, sensors, DEM,
atmospheric and topographic corrections methods.
In order to minimize topographic impacts, we
concentrated on topographic impacts on LC mapping
precision, in this study we suggested the
implementation of Modified Sun-Canopy-Sensor
Correction (SCSC) technique to generate LC maps
in Gua Musang district which is located in a
rugged mountainous terrain area in Kelantan
state, Malaysia.
3
Aim
This study aims to evaluate SCSC TC method on
the RBF-based SVM classification performance
using Landsat 8 imagery.
4
Materials and Methods
a- Study Area (Gua Musang) Gua Musang is biggest
district in Kelantan's state ,Malaysis (453'
3.4044''N and10158' 5.4408'' E) covering an area
4600 km2, with a broad altitude 1211770 m ASL.
This region accounts is mountainous terrain. The
primary LC is woodland, oil palm and rubber 14,
15.
b- Data Source Landsat 8 OLI (30 m) was used for
LC mapping on 22 April 2014, SPOT 5 image , ASTER
stereo-pair data
Location of Gua Musang, Malaysia
5
a-Data Pre-processing The pre-processing
includes 1-Convert Landsat 8 data DN values to
Reflectance by applying the spectral radiance
scaling factor followed by 2-Apply FLAASH
method to transform radiance values to the Top of
Atmospheric Correction (TOA). 3-Correct the
geometrics with 24 uniformly distributed GCPs,
then project it base on the SPOT 5 image register
using a bilinear conversion with an accuracy of
0.5 pixels. b- DEM FCC, NDVI, SAVI, SR
Generation
FCC image
NDVI image c- Apply SCSC
Topographic Correction (TC)
6
Methodology
d-RBF-Based SVM Classification Process it
involves 1- Propose classification system with
nine classes , 2- Collect 10538 pixels as
training data and 100 pixels as validation
samples for image classification process and
accuracy assessment using the ROI instruments
supplied by ENVI software (v. 5.1) 3- Assess the
quality of training data by using J M method
and the value was range from 1.7 to 2. 4- Apply
SVM base on RBF kernel and the parameters as
follow ? 0.167 , C 100 and pyramid
zero. e-Performance Accuracy Assessment The
accuracy done by error matrix and PA, UA, overall
accuracy (OA) and Kappa coefficient (k) measured
while the McNemar test is used for comparing
accuracies among classified images.
7
Methodology of Study
8
Results and Discussions
The comparison of UI and CI Landsat 8 OLI
imageries with RBF-based SVM classification
results are shown in the following two imagers
Uncorrected (TC) Corrected TC
Figure shows the topographic effects are clearly
seen in the rugged terrain in the UI. The slope
facing off the illumination source in the CI
appears brighter than the same path in the UI as
a consequence of TC
Figure UI and CI Landsat 8 OLI imagery with
RBF-based SVM classification results
9
Results and Discussions
From the table bellow , the OA for the Landsat 8
OLI imagery increase from 86.22 to 90.11, after
TC being performed. Forest classes have low
accuracies for the UI due to the topographic
effect because most of the forests are placed in
rugged terrain, which contributes to the
different illumination of slopes. As well as this
correction technique could enhance the accuracy
by 4. The McNemar test Z and P values measured
between CI image and UI about 6.42 and 0.0001
which mean the enhancement in the CI image is
statistically significant at p  0.05.
Uncorrected (TC)
Corrected TC
10
Results and Discussions
Land Cover Classes CI (with (SCSC) as Topographic Correction) CI (with (SCSC) as Topographic Correction) UI (without (SCSC) as Topographic Correction) UI (without (SCSC) as Topographic Correction)
Land Cover Classes PA UA PA . UA
Water Bodies 98 100 63 88.73
Forests 82 86.81 100 67.11
Oil Palm 89 85.44 96 88.07
Rubber 93 95.88 76 100
Crop Land 96 88.07 89 96.74
Grass Land 79 78.22 94 94
Barren Land 65 85.53 91 84.26
Built-Up Area 96 84.21 76 80
Others 100 98.04 100 100
OVA 90.11 86.22
KAPPA 0.88 0.85
11
Summary
The topographic impact may change the radiance
values captured by the spacecraft sensors
typically in rugged terrain landscapes, resulting
in distinct reflectance value for similar land
cover classes and mischaracterization. In order
to minimize topographic impacts, we suggested the
implementation of SCSC technique to generate
land cover maps in Gua Musang district ,Kelantan
state, Malaysia using an atmospherically
corrected Landsat 8 imagery by SVM algorithm. The
results showed that the SCSC method
significantly reduces the effect with improvement
about 4 classification accuracy. The McNemar
test shows the enhancement in the CI image is
statistically significant at p  0.05 with 6.42 Z
value.
12
References
1 A. Fahsi, T. Tsegaye, W. Tadesse, and T.
Coleman, "Incorporation of digital elevation
models with Landsat-TM data to improve land cover
classification accuracy," Forest Ecology and
Management, vol. 128, no. 1-2, pp. 57-64,
2000. 3 M. S. Rani, O. Schroth, R. Cameron, and
E. Lange, "The Effect of Topographic Correction
on SPOT6 Land Cover Classification in Water
Catchment Areas in Bandung Basin, Indonesia," in
GISRUK 2017 Proceedings, 2017, no. 96
Geographical Information Science Research
UK. 14M. Hossain, J. Bujang, M. Zakaria, and M.
Hashim, "Application of Landsat images to
seagrass areal cover change analysis for Lawas,
Terengganu and Kelantan of Malaysia," Continental
Shelf Research, vol. 110, pp. 124-148,
2015. 15B. Satyanarayana, K. A. Mohamad, I. F.
Idris, M.-L. Husain, and F. Dahdouh-Guebas,
"Assessment of mangrove vegetation based on
remote sensing and ground-truth measurements at
Tumpat, Kelantan Delta, East Coast of Peninsular
Malaysia," International Journal of Remote
Sensing, vol. 32, no. 6, pp. 1635-1650, 2011.
13
Thank You
Contact Information
Address Department of Civil Engineering, Faculty
of Engineering, Universiti Putra Malaysia 43400,
Serdang, Selangor, Malaysia Email
Aldoski-Jwan_at_hotmail.com, shattri_at_gmail.com,
zailanikhuzaimah_at_gmail.com
5-6/11/2019