The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
400
3) Select a certain number of individuals with high fitness value
as “elitism” of the population and retain them in the next
generation. It is in this way that the individuals with high fitness
value are retained in the population and the principle of
“survival of the fittest” of the Genetic Algorithm is conveyed.
4) Check whether the termination conditions are satisfied. If so,
the evolution stops and the optimal result represented by the
best individual is returned. Otherwise, the evolution continues
and the next generation is produced. The termination conditions
can be either a predefined fitness threshold or number of
generation evolved.
5) If the population continue to evolve, the next generation are
produced following procedure below. First, a certain number of
individuals are selected randomly to compete the mating right.
Two individuals of the highest fitness values are selected as a
pair of parents. Crossover is operated on their chromosomes to
produce two children individuals. Location of the crossover
point on the chromosomes is also randomly determined; Second,
a floating-point number in the range of 0.0 to 1.0 is generated
randomly. If it is less than the predefined mutation possibility,
mutation is operated for the two children individuals; Repeat
the two steps before to produce all the children individuals
(except the “elitism” individuals ) in the new generation.
6) Repeat operations from step 2) to step 4).
2.3.4 Realization of the GA-SVM
The GA-SVM method was realized in the ENVI/IDL language.
IDL (Interactive Data Language) is an array-oriented,
interpreted, structured programming language that provides
powerful data analysis and visualization capabilities. ENVI
(Environment for Visualizing Images) is a software for the
visualization, analysis, and presentation of all types of digital
imagery. It is written in IDL, so most of its functionalities are
provided to users in the form of functions that can be easily
called in programming.
In our research, we realized functionalities like opening and
closing of the hyperspectral images, extraction data in the
training ROI (Regions of Interest) and testing ROI, performing
classification using the SVM, as well as evaluation of
classification accuracy all through calling existing function
provided by ENVI. The Genetic Algorithm was realized using
the Object-Oriented programming technique of IDL. Two
object classes, i.e. population and chromosome, were created
with their properties defined respectively. Methods were then
defined for these two object classes, respectively. Methods for
the population class include population initialization,
tournament selection, chromosome crossover and mutation etc.
Methods for the chromosome class include coding and decoding
of the chromosomes, setting and getting of chromosome
properties etc. Finally, the GA-SVM method was programmed
according to the steps described in section 2.3.3.
3. EXPERIMENTS
3.1 Data Sets
The hyperspectral data used in this study is a cloudless
Hyperion image taken on Dec. 18, 2005. The image centers at
113° 20 ' 48 " E, 23° 5 ' 36 " N and covers part of the
Guangzhou city, China.
The Hyperion system on board the EO-1 platform provides a
high resolution hyperspectral imager capable of resolving 220
spectral bands (from 0.4 to 2.5 pm) with a 30-meter resolution.
The instrument can image a 7.5 km by 100 km land area per
image, and provide detailed spectral mapping across all 220
channels with high radiometric accuracy. The Hyperion images
has wide ranging applications in mining, geology, forestry,
agriculture, and environmental management. Detailed
classification of land assets through the Hyperion will enable
more accurate remote mineral exploration, better predictions of
crop yield and assessments, and better containment mapping.
In this study, we selected part of a Hyperion image that covers
the Haizhu district of Guangzhou city to test the GA-SVM
method. Atmospheric correction and geometric correction of
the image were first performed in order to eliminate
atmospheric effects and compare to the ground truth data.
3.2 Feature selection for the Hyperion image using the GA-
SVM method
The steps of feature selection for the Hyperion image using the
GA-SVM method include:
1) Create the training sets and testing sets needed by the GA-
SVM method using the ENVI software. Land uses of the study
area were categorized into six classes, namely built-up area,
water body, grassland, forest and unused land. Training sets and
testing sets were created for each land use class.
2) Set parameters of the GA-SVM method. These parameters
include: range of kernel parameter C and , bit lengths for the
three chromosome parts, initial population size, population size
of each generation, elitism size, number of generation to evolve,
tournament size, crossover rate and mutation rate. These
parameters were set as follow: min_C = 90.0 , max_C
=.110.0 , miny = 0.9/bit length , maxy = 0.25 ,
initPopulationSize = 100 , populationSize - 50 »
numGenerations = 40, offspringPerGen = 46. tournamentSize
= 6, crossOverRate = 0.98, elitism = 4, mutationRate = 0.02.
3) Run the GA-SVM model with the inputs of Hyperion image,
training sets and testing sets.
c
Y
Selected Bands (with their central
wavelength, unit: mm)
Classification
Accuracy
Kappa
Coefficient
No Feature Selection
Feature Selection
100.000
0.000
All 196 bands
4(457.34), 12(538.74), 25(671.02),
33(752.43), 61(1033.50), 67(1094.09),
88.81%
0.8619
using the GA-SVM
Method
95.0297
0.2021
75(1174.77), 88(1305.96), 106(1487.53),
109(1517.83), 117(1598.51), 127(1699.40),
132(1749.79)
92.51%
0.9261
Table 1. Performance of the GA-SVM Feature Selection