Denoising and Wavelet-Based Feature Extraction
of MODIS Multi-Temporal Vegetation Signatures
Lori Mann Bruce and Abhinav Mathur
Electrical and Computer Engineering Department
Mississippi State University
Outline
• Project Goals
• MODIS Data
• Denoising Methods
• Feature Extraction Methods
• Experimental Results
• Conclusions
MODIS Data For Invasives Detection
Time
ND
VI
Target VegetationAlternate Vegetation
Noise in Spectral Signatures• Encountering problems with Quality Assurance (QA) of MODIS
imagery
Hierarchical Data Format (HDF) – Self describing file format
Science Data Sets (SDSs) – 2D, 3D or 4D arrays
Attributes – text or other data that annotates the file (global) or arrays (SDSs)
Metadata – ECS metadata for products (stored as attributes)
.met file contains the ECS core metadata (includes QA information, date/time products acquired/produced, etc.)HDF-EOS Metadata - SWATH or GRID – (includes geometric information that relates data to specific earth locations)
Bit No. Parameter Name Bit Comb.
Description
00 VI produced with good quality 01 VI produced but with unreliable quality and thus examination of other
QA bits recommended 10 VI produced but contaminated with clouds
0-1 VI Quality (MODLAND Mandatory QA Bits)
11 VI not produced due to bad quality 0000 Perfect quality (equal to VI quality = 00: VI produced with good
quality) 0001 High quality 0010 Good quality 0011 Acceptable quality 0100 Fair quality 0101 Intermediate quality 0110 Below intermediate quality 0111 Average quality 1000 Below average quality 1001 Questionable quality 1010 Above marginal quality 1011 Marginal quality 1100 Low quality 1101 No atmospheric correction performed 1110 Quality too low to be useful
2-5 VI Usefulness Index
1111 Not useful for other reasons (equal to VI quality = 11: VI not produced due to bad quality)
00 Climatology used for atmospheric correction 01 Low 10 Intermediate
6-7 Aerosol Quantity
11 High 0 (No) No adjacency correction performed 8 Atmosphere
Adjacency Correction
1 (Yes) Adjacency correction performed
0 (No) No atmosphere-surface BRDF coupled correction performed 9 Atmosphere BRDF Correction 1 (Yes) Atmosphere-surface BRDF coupled correction performed
0 (No) No mixed clouds 10 Mixed Clouds 1 (Yes) Possible existence of mixed clouds 00 Ocean/inland water
Shallow ocean Moderate and continental ocean Deep ocean Deep inland water
01 Coastal region Ocean coastlines and lake shorelines Shallow inland water
10 Wetland Ephemeral water
11-12 Land/Water Mask
11 Land 0 (No) No snow/ice 13 Snow/Ice 1 (Yes) Possible existence of snow/ice 0 (No) No shadow 14 Shadow 1 (Yes) Possible existence of shadow 0 BRDF composite method used for compositing 15 Compositing Method 1 Constraint view angle MVC (CV-MVC) method used for compositing
MODIS images from January 2001 to December 2003
Months: 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12
Click on the image
Time line
Time
EV
I val
ue
MODIS images from January 2001 to December 2003
Denoising MODIS Time-Series Data
veg type1
veg type2
moving average filter median filter
0 10 20 30 40 50 60 70
0 10 20 30 40 50 60 70
0 10 20 30 40 50 60 70
0 10 20 30 40 50 60 70
0 10 20 30 40 50 60 70
0 10 20 30 40 50 60 70
Feature Extraction from MODIS Time-Series Data
Fourier Analysis
phase response
0 100 200 300-4-2024
0 100 200 300-4-2024 phase response
0 100 200 3000
1
2
3magnitude response
0 100 200 3000
1
2
3magnitude response
0 10 20 30 40 50 60 70
denoised veg type1
0 10 20 30 40 50 60 70
deniosed veg type2
Fourier-Based Feature Extraction
10 20 30 40 50 60
0.51
1.5
22.5
Ma
gn
itude
Frequency sample points10 20 30 40 50 60
-2
0
2
Ph
ase
F1
F2 F4
F3
Wavelet Decompositions
Highpass & Lowpass Decomposition FiltersCorresponding to Selected Mother Wavelet
LP 2 HP 2
2LP
),2(, xW Nf
f(x)HP
HP 2
2
.
.
....
LP 2LP 2...
),2( 1, xW Nf
),2( 2, xW f
),2( 1, xW f
k
kjfj
xWxfkj
)()( ,, ,
)(),( ,, ,xxfW kjf kj
Inverse DWT
Discrete Wavelet Transform (DWT)
Wavelet-Based Feature Extraction
Temporal Signature Haar Mother Wavelet
Approximation CoefficientsSignal Approximation
Scale 2^3
Scale 2^2
Scale 2^1
Wavelet-Based Feature Extraction
Mean
F1, F2, …, F6
Fourier-Based Features
Veg1 Noisy
Veg2 Noisy
Veg1 Denoised
Veg2 Denoised
Mean F1 2.63e05 2.79e05 2.63e05 2.78e05 Std F1 1.9e08 1.07e08 1.89e08 1.06e08
Mean F2 0 0 0 0 Std F2 0 0 0 0
Mean F3 1.26e05 1.23e05 1.23e05 1.20e05 Std F3 3.34e07 4.46e07 3.21e07 4.27e07
Mean F4 -3 -3 1 -1 Std F4 0 0 9 10
Wavelet-Based Features
Veg1Noisy
Veg2Noisy
Veg1Denoised
Veg2Denoised
Mean F1 834 1654 1339 1618
Std F1 109438 134644 111950 108233
Mean F2 3424 4662 2092 3820
Std F2 108333 143233 95250 195662
Mean F3 11001 10184 9151 8454
Std F3 578968 706544 423224 447922
Mean F4 15200 15251 14785 4612
Std F4 179830 92016 183090 74526
Mean F5 9902 11302 11641 12697
Std F5 784840 697503 671643 695616
Mean F6 3677 4159 5496 6193
Std F6 190165 102728 251508 118408
Classification Accuracies
OverallVeg2Veg1
67%56%75%Denoised – NN
81%89%75%Denoised – ML
67%56%75%Noisy – NN
81%89%75%Noisy – ML
100%100%100%Denoised – NN
100%100%100%Denoised – ML
95%100%92%Noisy – NN
95%100%92%Noisy – ML
OverallVeg2Veg1
Fourier-BasedFeatures
Wavelet-BasedFeatures
Conclusions
• MODIS time-series data has isolated noise spikes
• Fourier-based features less affected by noise than wavelet-based features
• Shape-preserving features needed for invasives detection project
• Wavelet-based features resulted in significantly higher accuracies than Fourier-based features
• Simple denoising methods (moving average or median filter) were sufficient
Wavelet Analysis
10 20 30 40 50 60 70
0
2000
4000
6000
8000
a1
2000
4000
6000
a2
1000
2000
3000
4000
5000
6000
a3
0
2000
4000
6000
8000
s
Signal and Approximation(s)
cfs
Coefs
10 20 30 40 50 60 70
3
2
1
10 20 30 40 50 60 70
2000
4000
6000
8000
a1
2000
4000
6000
a2
2000
3000
4000
5000
6000
a3
2000
4000
6000
8000
s
Signal and Approximation(s)
cfs
Coefs
10 20 30 40 50 60 70
3
2
1
0 10 20 30 40 50 60 70
denoised veg type1
0 10 20 30 40 50 60 70
deniosed veg type2
Feature Extraction from MODIS Time-Series Data
