Effect of Data Transformation Methods on SVM Performance for Land Cover Mapping

1
Fifth International Scientific Research
conference Renewable Energy Water
Sustainability, Tanta University, Egypt, March
26-28, 2019
Effect of Data Transformation Methods on SVM
Performance for Land Cover Mapping

By Jwan Aldoski
Prof. Shattri B. Mansor Dr. Zailani Khuzaimah
Universiti Putra Malaysia, Faculty of
Engineering, Department of Civil Engineering,
Address Jalan UPM ,43400, Serdang,
Selangor, Malaysia Emails
Aldoski-Jwan_at_hotmail.com,
shattri_at_gmail.com,
zailanikhuzaimah_at_gmail.com
2
Presentation Outline 
3
Introduction

• Land Cover (LC)
• The Earths cover surface information mainly
  provided by Remotely Sensed Images through using
  parametric and non-parametric classification
  algorithms.
• To overcome parametric algorithms problem some
  non-parametric classification techniques have
  been introduced like Support vector machine (SVM)
  .
• SVM is supervised learning method used recently
  for LC mapping .
• The ability of SVM is affected by features type
  that extracted from remote sensing data, thus,
  transformation feature derived from remote
  sensing data maybe performed to achieve the best
  LC mapping result for a given data set .
• To choice transformation techniques for extracted
  transform features required optimal results.

4
The Aims
The objective of this study is twofold (1)
Evaluate the potential of Landsat 8 imagery and
SVM classification method for LC mapping. (2)
Evaluate the effects of four image transformation
techniques include Principal Components Analysis
(PCA), Independent Components Analysis (ICA) ,
Minimum Noise Fraction Transform (MNF) and
Tasselled Cap Transformation (TCT) on the SVM
classification accuracy
5
Study area
A

• The test area is the Kota Bharu district ,
  Kelantan sate, Malaysia.
• The precipitation is more than 6,000 mm of mean
  rainfall annually.
• Annual temperature is about 27.5 C.
• Oil palm is the major cash tree in the region.
• The elevation increases gradually from 400 - 900
  m above sea level
• Study area is relatively flat

B
C
Malaysia
Kota Bharu
Figure 1. (A) The research area over the Kota
Bharu district (imposed over a natural-color
composite of Landsat-8 image with multispectral
band (RGB 4, 3, 2). (B) Blowup of the study
area (C) Schematic map of the research area,
Malaysia borders, and the Landsat 8 image
footprint.
6
Methodology
A preprocessing Landsat 8 OLI Data May 19, 2017
1- Geometric correction using 15 GCPs and
(DEM) with (RMSE) 0.07 2- Atmospheric
correction using FLAASH algorithm 3- Band
selection 2, 3, 4, 5, 6, and 7 bands 4-
Sub-set and Re-projection process
Figure 2. A sample of false color composites of
landsat8 imagery in the study area in the Kota
Bharu district. RGB correspond to band 5 near
IR (0.850.88 lm), band 4 red (0.640.67 lm),
and band 3 green (0.530.59 lm)
7
Methodology

• B-Derived Transformation Image
• Transformation techniques include
• Principal Components Analysis (PCA),
• Independent Components Analysis (ICA),
• Minimum Noise Fraction Transform (MNF), and
• TCT

Figure 6. A sample of false color composites of
TCB, TCG, and TCW landsat8 imagery
Figure 4. A sample of false color composites of
PC1, PC2, and PC3 landsat8 imagery
8
Methodology
C- SVM Implementation and post processing

1. Collecting Training and test data using Random
   Pixel Selection Strategy

(2) Spectral separability
Separability Values
LULC Categories LULC Categories 1 2 3 4 5 6 7 8 9
1 Water Bodies 1 1.97 1.99 1.98 1.99 1.96 1.99 1.96 1.8
2 Urban Area   1 1.99 1.99 1.99 1.99 1.95 1.94 1.61
3 Primary Forest     1 1.62 1.7 1.66 1.89 1.92 1.99
4 Secondary Forest       1 1.55 1.58 1.78 1.85 1.99
5 Oil Palm         1 1.54 1.98 1.95 1.99
6 Rubber           1 1.99 1.97 1.99
7 Agricultural Fields             1 1.85 1.97
8 Grass Area               1 1.96
9 Barren Area                 1
9
Methodology
C- SVM Implementation and post processing
SVM parameters RBF kernel function Penalty
value C120 Kernel parameter ? 0.15 probability
threshold value 1
(3) Apply SVM

• Five classification strategies were designed
• only the bands 27 from Landsat 8 (6 bands)
• PCA1-3, (3 bands),
• 1-6 ICA (6 bands),
• 1-5 MNF (5 bands) and
• 3 TCT (3 bands).

( 4) Post Classification Processing using
Majority Filtering
Parameters windows size 33 pixels by pixels with
centre pixel weight (1 value)
(5) Accuracy Assessment using Error Matrices
( 6) statistical significance using Z-test
10
Results

• LC maps based on spectral and transformation
  features derived from Landsat 8 image were
  evaluated using
• Visual Evaluation
• Error Matrices
• Area Evaluation
• Statistical Significance

11
Results

• Visual Evaluation

(b) PCA
(a) Original 6 Landsat 8 OLI bands

• PCA map detect water very well
• Original 6 Landsat 8 OLI bands Map detect the
  agricultural field
• Original 6 Landsat 8 OLI bands, PCA and ICA
  detect oil palm, rubber, and agricultural fields
  very well
• PCA and ICA maps detect barren area and urban
  area well
• While distinguish between primary forest and
  secondary forest based on the PCA and ICA almost
  impossible

12
Results

• Visual Evaluation

(e) TCT
(c) ICA
(d) MNF
13
Results

• Error Matrices

• The overall trend of LC mapping accuracy ranked
  in terms of OA and kappa statistics, are as
  follows
• ICA gt PCA gt original 6 bands gt MNF gt TCT
• The highest LC maps accuracy could be attributed
  to the fact that PCA extended the possibility for
  pattern recognition since the imagery data were
  transformed into new, uncorrelated co-ordinate
  system or vector space. One other hand, ICA as
  expected increase accuracy since it maximized the
  between classes variance and minimized the within
  class variance, this in turn let to an increase
  separability between classes.

  Original 6 bands PCA ICA MNF TCT
Overall Accuracy () 81.50 89.50 90.28 80.33 74.78
Kappa Coefficient (?) 0.79 0.88 0.89 0.78 0.72
14
Results

• Error Matrices

15
Results

• Error Matrices

16
Results

• Area Evaluation

• Agricultural fields and oil palm covered greatest
  portion about followed by rubber and secondary
  forest about (10and 8 separately) as second
  major land cover.
• Forest covering in the southern parts of the
  area were mainly on hilly areas.
• From here we conclude that selection of
  transformation methods not only affects the
  accuracy of the map and also when calculating the
  area they have a clear effect and accordingly
  affects the decision of land management.

17
Results

• Statistical Significance

SVM-classifier base 1 SVM-classifier base 2 Comparison of Overall accuracy(OV) Comparison of Overall accuracy(OV) Comparison of McNemar test pair-wise (95 confidence level) Comparison of McNemar test pair-wise (95 confidence level) Comparison of McNemar test pair-wise (95 confidence level)
SVM-classifier base 1 SVM-classifier base 2 OV1 OV2 z Observed value z Critical value p Value
original 6 bands PCA 81.5 89.5 12.37 1.96 lt0.0001
original 6 bands ICA 81.5 90.28 11.52 1.96 lt0.0001
original 6 bands MNF 81.5 80.33 1.82 1.96 0.069
original 6 bands TCT 81.5 74.78 7.97 1.96 lt0.0001
PCA ICA 89.5 90.28 0.73 1.96 0.466
PCA MNF 89.5 80.33 12.22 1.96 lt0.0001
PCA TCT 89.5 74.78 15.45 1.96 lt0.0001
ICA MNF 90.28 80.33 12.02 1.96 lt0.0001
ICA TCT 90.28 74.78 15.09 1.96 lt0.0001
MNF TCT 80.33 74.78 7.54 1.96 lt0.0001
SVM with PCA resulted 0.8 increase in
classification accuracy than the SVMs with ICA
which was statistically insignificant or
statistically no difference between
classification accuracy.
18
Conclusions

• This study was conducted to determine the main LC
  classes using different data transformation
  methods with Landsat-8 OLI image.
• PCA, ICA, MNF and TCA were used to increase the
  spectral dissimilarities between LC classes in
  order to improve SVM classification accuracy.
• The dominant LC classes in the study area were
  agricultural fields and oil palm.
• PCA, MNF and ICA Data transformation improved the
  accuracy of SVM classification. ICA significantly
  improved the classification accuracy with OA
  90.28, followed by PCA with OA 89.5. On the
  other hand, the TCT components did not improve
  the classification accuracy.
• The Z test shows ICA not statistically
  significant differences with SVM base on PCA
  given the z value 0.73
• Primary Forest And Secondary Forest Classes as
  they consist of dense tropical trees were showed
  different produces and user accuracy values using
  original 6 Landsat 8 bands and transformation
  images need more study and test

19
Thank You
Any questions?