06 Feature Classification Using Supervised

Therese Anne M. Rollan
PARMap
Introduction
pinkhawk.hubpages.com/hub/coffee-production-philippines
www.commercialpressuresonland.org/press/da-wants-ecozones-created-agribusiness

Agriculture is very
important in the economy
of the Philippines with its
goal of reducing poverty
and hunger by 2015 from
its level from 1990 (Albert,
2013).
www.gfreefoodie.com/wp-content/uploads/2013/03/golden-rice.jpg
www.pepper.ph/is-our-food-safe-philippines-on-guard-as-gmo-wheat-found-in-us-imports/
Agricultural Mapping

predicting grain crop yield,

collecting crop production statistics,

facilitating crop rotation records,

mapping soil productivity,

identification of factors influencing crop stress, and

assessment of crop damage and monitoring farming activity.
Objectives
1
Principles
Behind
SVM
2
Resources and
Derivable Information
3
Classification
Workflow
4
Tests
5
Recommendations
SVM
Support Vector Machines (SVM) Introductory Overview. Retrieved from http://www.statsoft.com/Textbook/Support-Vector-Machines
Introduction to Support Vector Machines. Retrieved from http://docs.opencv.org/doc/tutorials/ml/introduction_to_svm/introduction_to_svm.html
SVM
Optimal
Hyperplane
Support Vector Machines (SVM) Introductory Overview. Retrieved from http://www.statsoft.com/Textbook/Support-Vector-Machines
Introduction to Support Vector Machines. Retrieved from http://docs.opencv.org/doc/tutorials/ml/introduction_to_svm/introduction_to_svm.html
SVM
Support
Vectors
Optimal
Hyperplane
Support Vector Machines (SVM) Introductory Overview. Retrieved from http://www.statsoft.com/Textbook/Support-Vector-Machines
Introduction to Support Vector Machines. Retrieved from http://docs.opencv.org/doc/tutorials/ml/introduction_to_svm/introduction_to_svm.html
Kernel Functions
Optimal
Hyperplane
Nikhil Garg. What are Kernels in Machine Learning and SVM? Retrieved from http://www.quora.com/What-are-Kernels-in-Machine-Learning-and-SVM
KERNEL FUNCTIONS
Kernel
Function
Descriptions
Linear
 Performs well if the number of features is large compared to the size
of the data, and if is unnecessary to map to a higher dimensional space
 Deals poorly with noisy data
Polynomial
 Has the most tunable parameters of all the kernels
 Higher degrees tend to overfit, values may result to zero or infinity
Radial
Basis
Function
 Maps samples into the higher dimensional space in a non-linear fashion
 Not suitable when number of features is ”very large”
 The best kernel to ”start with”
Sigmoid
 Machine commonly found in neural networks
 Not valid for some parameter choices.
Rita McCue. A Comparison of the Accuracy of Support Vector Machine and Naive Bayes Algorithms In Spam Classification.
Retrieved from http://classes.soe.ucsc.edu/cmps242/Fall09/proj/RitaMcCueReport.pdf
Related Study
Supervised Parametric Classification of Aerial LiDAR Data
Amin P. Charaniya, Roberto Manduchi, and Suresh K. Lodha
Variables:
 Multiple Returns
 Luminance
 Intensity
 Normalized Height
 Height Variation
Amin P. Charaniya, Roberto Manduchi, and Suresh K. Lodha. Supervised Parametric Classification of Aerial LiDAR Data.
Retrieved from http://users.soe.ucsc.edu/~manduchi/papers/Amin3D.pdf
Related Study
Related Study
Amin P. Charaniya, Roberto Manduchi, and Suresh K. Lodha. Supervised Parametric Classification of Aerial LiDAR Data.
Retrieved from http://users.soe.ucsc.edu/~manduchi/papers/Amin3D.pdf
Classification Workflow
General Process in Feature Classification using SVM
Classification
mentalfloss.com/article/49774/dont-changething-8-inventions-never-needed-updating
Pre-processing
 Obtaining
Variables
Accuracy Assessment
missmernagh.com/2013/08/31/junior-infant-maths-sorting-classifying/
An
ITERATIVE
PROCESS
sr.photos3.fotosearch.com/bthumb/CSP/
CSP992/k12628895.jpg
Note: Variables and Training pixels per Class derived from
Amin P. Charaniya, Roberto Manduchi, and Suresh K. Lodha. Supervised Parametric Classification of Aerial LiDAR Data.
Retrieved from http://users.soe.ucsc.edu/~manduchi/papers/Amin3D.pdf
Classification
Workflow
Pre-processing
 Obtaining
Variables
Classification
Accuracy
Assessment
Classification
Workflow
Pre-processing
 Obtaining
Variables
Classification
Workflow
Classification
Accuracy
Assessment
Resources used

Data
LiDAR point cloud files (.las)
Orthophotos

Software
Las Tools
Envi
Testing Implementations
Testing Implementations
Classes/ROIs
Bare
Land
Fallow
Road
Built-up
Rice
Mango
Sugar
Trees
Water
Lasmerge
Lasheight
LiDAR Parameters
Texture and Layer Stacking Parameters
SVM Parameters
Variables
Intensity
Normalized
Height
Normalized
Height
Variance
Variables
HLS
Luminance
Variables
Green-Red Vegetation Index (GRVI)

Green vegetation (conifers,
deciduous trees, and grass):
green is higher than red.

Soils (brown sand, silt, and
dry clay): green is lower than
red.

Water/snow: green and red
are mostly the same.
Takeshi Motohka, Kenlo Nishida Nasahara, Hiroyuki Oguma and Satoshi Tsuchida.
Applicability of Green-Red Vegetation Index for Remote Sensing of Vegetation Phenology.
Retrieved from http://www.mdpi.com/2072-4292/2/10/2369
Testing Implementations
Summary of Parameters and Variables
LiDAR Parameters
SVM Parameters
Kernel Type
Gamma
Penalty
Parameter
Radial Basis Function
0.063
100
Model Type Interpolation Data Type
Full Feature
Linear
Double
Variables
Trial
1
2
3
4
5
6
7
I
IH
NH
NHV
L
RGB
GRVI
Red
Textures
Testing Implementations
Accuracy Assessment
Trial
Overall
Acc.
Kappa
Coeff.
1
2
3
4
5
6
7
71.53%
80.66%
74.00%
80.20%
81.27%
80.93%
85.27%
0.6734
0.7794
0.7021
0.7742
0.7863
0.7825
0.8319
Variables
I
IH
NH
NHV
L
RGB
GRVI
Red
Textures
Trial 1: I, NH & NHV
Testing Implementations
Testing Implementations
Trial 7: I, NH, NHV, L, RGB and Red Textures
TRIAL 1
TRIAL 7
TRIAL 1
TRIAL 7
TRIAL 1
TRIAL 7
TRIAL 1
TRIAL 7
Orthophoto
LiDAR
Orthophoto
+
LiDAR
Summary and Conclusion

The testing implementations used LiDAR point cloud data and Orthophotos
for the feature classification of a 2 by 2 km Tarlac dataset.

The test area contains 3 types of high-value crops namely Sugarcane, Rice
and Mango. All in all there are 9 classes identified with the remaining 6
specifically Trees (non-crop), Road, Built-up, Fallow, Bare Land and Water.

7 trials were done with Trial 7 as the most accurate classification having
85.27% overall accuracy and 0.83 kappa statistic.

Intensity Homogeneity does not provide significant contribution in
classification.

Using only 1 derivative from RGB is enough.

The combination of LiDAR and Orthophotos data yields higher accuracy than
using these datasets individually.
Recommendations

Experiment on other LiDAR Derivatives.

Explore different combinations of LiDAR and SVM parameters .

Try the workflow on larger datasets.

Every trials must be well-documented.

If available, use high-resolution satellite images with NIR band in order to
apply vegetation indices.
References

Albert, J. R. (2013, April 12). Beyond the Numbers: How Important is Agriculture in the Economy? Retrieved September 30,
2014, from National Statistics Coordination Board:
http://www.nscb.gov.ph/beyondthenumbers/2013/04122013_jrga_agri.asp#f1

Support Vector Machines (SVM) Introductory Overview. Retrieved from http://www.statsoft.com/Textbook/Support-VectorMachines

Introduction to Support Vector Machines. Retrieved from
http://docs.opencv.org/doc/tutorials/ml/introduction_to_svm/introduction_to_svm.html

Support Vector Machine Background. Retrieved from http://www.exelisvis.com/docs/BackgroundSVM.html

Nikhil Garg. What are Kernels in Machine Learning and SVM? Retrieved from http://www.quora.com/What-are-Kernels-inMachine-Learning-and-SVM

RBF SVM parameters. Retrieved from http://scikit-learn.org/stable/auto_examples/svm/plot_rbf_parameters.html

Rita McCue. A Comparison of the Accuracy of Support Vector Machine and Naive Bayes Algorithms In Spam Classification.
Retrieved from http://classes.soe.ucsc.edu/cmps242/Fall09/proj/RitaMcCueReport.pdf

Amin P. Charaniya, Roberto Manduchi, and Suresh K. Lodha. Supervised Parametric Classification of Aerial LiDAR Data.
Retrieved from http://users.soe.ucsc.edu/~manduchi/papers/Amin3D.pdf

Lodha, Suresh K., Kreps, Edward J., Helmbold, David P., and Fitzpatrick, Darren. Aerial LiDAR Data Classification using
Support Vector Machines (SVM). Retrieved from http://users.soe.ucsc.edu/~dph/mypubs/LidarSVM.pdf