Modular Neural Networks Modular Neural Networks Approach to Chemical Approach to Chemical Content Analysis of Content Analysis of Vegetation Vegetation 1 N. Kussul, N. Kussul, 1 V. Yatsenko, V. Yatsenko, 2 A. Sachenko, A. Sachenko, 3 G. G. Markowsky, Markowsky, 1 A. Sydorenko, A. Sydorenko, 1 S. Skakun, S. Skakun, 2 S. Ganzha S. Ganzha 1 Space Research Institute NASU- NSAU, 40 Glushkov Ave 03187 Kiev, Ukraine, [email protected]v.ua 2 Institute of Computer Information Technologies of Ternopil Academy of National Economy 3 Peremoga Square, 46004, Ternopil, Ukraine, [email protected]3 Department of Computer Science, 5752 Neville Hall, University of Maine, Orono, ME 04469- 5752, [email protected]
21
Embed
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation 1 N. Kussul, 1 V. Yatsenko, 2 A. Sachenko, 3 G. Markowsky, 1 A. Sydorenko,
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Modular Neural Networks Modular Neural Networks Approach to Chemical Content Approach to Chemical Content
Analysis of VegetationAnalysis of Vegetation
11N. Kussul, N. Kussul, 11V. Yatsenko, V. Yatsenko, 22A. Sachenko, A. Sachenko, 33G. Markowsky,G. Markowsky,11A. Sydorenko, A. Sydorenko, 11S. Skakun, S. Skakun, 22S. GanzhaS. Ganzha
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
Introduction
Contents . . .
Architecture
Problem solution
Experimental results
Comparison
Conclusions
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
Introduction
Contents . . .
Architecture
Problem solution
Experimental results
Comparison
Conclusions
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
IntroductionSpectral characteristics of light, which is reflected from Earth objects, represent convenient and high informative data sources for remote investigations. It can be used for estimation of vegetation state to determine infection and pollution level of vegetation.
150
250
350
450
550
400 600
2,76
6,82
6,15
9,18
4,62
3,75
7,65
1,46
5,29
Intensity dependence of reflected light on wave-length with different chlorophyll content
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
Each spectral curve contains 350 points, which determines the dimension of Neural Network input layer. It is evident that high dimension of input data and large training set requires the use of modular Neural Network architecture.
150
250
350
450
550
400 600
2,76
6,82
6,15
9,18
4,62
3,75
7,65
1,46
5,29
Intensity dependence of reflected light on wave-length with different chlorophyll content
Introduction
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
Introduction
Contents . . .
Architecture
Problem solution
Experimental results
Comparison
Conclusions
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
To determine plants damage (infection) level a modular Neural Network is used. It consists of classifier and interpolator.
Classifier InterpolatorXInput
Y
X
(1-Y)XY=0 - damagedY=1 - undamaged
Classifier InterpolatorXInput
X
(1-Y)XY=0 - damagedY=1 - undamaged
Chlorophyllcontent
Architecture
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
Classifier executes data pre-processing (brute classification), dividing input data into 2 classes: damaged and undamaged.
Classifier InterpolatorXInput
Y
X
(1-Y)XY=0 - damagedY=1 - undamaged
Classifier InterpolatorXInput
X
(1-Y)XY=0 - damagedY=1 - undamaged
Chlorophyllcontent
Architecture
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
If classifier output is 0 (i.e. input pattern is classified as damaged), then it is put on interpolator input.
Classifier InterpolatorXInput
Y
X
(1-Y)XY=0 - damagedY=1 - undamaged
Classifier InterpolatorXInput
X
(1-Y)XY=0 - damagedY=1 - undamaged
Chlorophyllcontent
Architecture
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
Introduction
Contents . . .
Architecture
Problem solution
Experimental results
Comparison
Conclusions
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
Before the investigation of modular architecture effectiveness is done, we will define the best training parameters of Neural Network and find the quantitative rates of training process
Problem solution
Training Methods Number of training epochs
Fletcher-Powell method not trained
Levenberg-Marquardt method 456 168 117
Back propagation on-line 2740 2644 2407
Back propagation off-line not trained
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
To estimate the best Neural Network training parameters appropriate experiments were run.
0
100000
200000
0 0,05 0,1 0,15 0,2 0,25 0,3 0,35
Dependence of number of training epochs on learning coefficient (full range)
Problem solution
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
Dependence of number of training epochs on learning coefficient (smaller range)
2500
3500
4500
5500
0 0,05 0,1
It is evident that the best values are the following: learning coefficient — 0.06, moment coefficient — 0.125.
Problem solution
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
It is evident that the best values are the following: learning coefficient — 0.06, moment coefficient — 0.125.
0
10000
20000
30000
0 0,1 0,2 0,3 0,4 0,5 0,6
Dependence of number of training epochs on moment coefficient
Problem solution
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
Introduction
Contents . . .
Architecture
Problem solution
Experimental results
Comparison
Conclusions
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
Obtained experimental results showed that both types of classifiers train quickly enough (classifier of the first type for 300-400 epochs, and classifier of the second type — for about 20 epochs.
Classifier training process
Experimental results
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
For interpolator a described above multi-layered Neural Network was used. A training set has smaller dimension.
0
200
400
600
800
1000
1200
0 0,02 0,04 0,06 0,08 0,1 0,12
0
100
200
300
400
500
600
0 0,1 0,2 0,3 0,4 0,5 0,6
Dependence of interpolator training time on learning
coefficient
Dependence of interpolator training time on moment
coefficient
Experimental results
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
Introduction
Contents . . .
Architecture
Problem solution
Experimental results
Comparison
Conclusions
Forward
Back
Modular Neural Networks Approach to Chemical Content Analysis of Vegetation
Conducted experiments showed that modular architecture has advantages over traditional in the sense of training time.
Comparative training time analysis of traditional and modular NN architectures. On x-axis there are values of learning coefficients (uniform fill) and moment coefficients (line fill). On y-axis there is a ratio between numbers
of training iterations for traditional NN (T) and for modular NN (M)