Implementation of Local Maximum Fitting in Graphic

Improvement of Local Maximum Fitting (LMF)
for High Temporal Remote Sensing Data
Using Meteorological Data
_________________________________________________________________________________________________________________________________________________________________
Examination Committee:
External committee:
Dr. Kiyoshi Honda (Chairperson)
Prof. Seishiro Kibe (Member)
Dr.Vivarad Phonekeo (Member)
Mr. Wataru Ohira
By Salinthip Kungvalchokechai
RS & GIS FoS, SET
1
Contents
_______________________________________________________________________________________________________________________________________
1.
2.
3.
4.
Introduction
 Background: Local Maximum Fitting (LMF)
 Statement of the Problems
 Objective
Methodology
Results
Conclusion & Recommendation
2
Introduction
_____________________________________________________________________________________________________________________________ __________


The seasonal vegetation changes are monitored by high frequent
satellite observation.
Original satellite data is contaminated of noise by cloud and
haze, especially in rainy season, which degrade an original
satellite NDVI data.
Local Maximum Fitting (LMF) is developed to remove the
effect of cloud and haze. (Sawada, H., 2001)
Comparison of Original NDVI and after LMF (WIN=3) in 2005
2000
1900
1800
1700
NDVI

1600
1500
1400
1300
1200
1100
1000
Jan
Feb
Mar
Dry Season
Apr
May
Jun
Aug
Jul
Rainy Season
Sep
Oct
No
Dec
Dry
Many fluctuate noise
in rainy season
Month
Original NDVI
after LMF (WIN=3)
3
Background: LMF
______________________________________________________________________________________________________________________________________
3 Step of LMF: Local Maximum Fitting
 Min Max Filter
d 't  Min[ Max(dt  w1 , dt  w 2 ,
, dt ), Max(dt , dt 1 ,
, dt  w1 )]
dt : observed data at time t
w : filter window
d 't : modified data at time t

Window filter size = the number of window which
we use to filter noise at the considering point.
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000
NDVI after Min Max Filter (WIN=3) from 2005 to 2007
NDVI
NDVI
Original NDVI from 2005 to 2007
Jan
Feb Mar Apr May
Dry Season
Jun
Jul Aug Sep
Rainy Season
Oct No
Dec Jan Feb Mar
Dry Season
Apr May Jun
Jul
Aug Sep
Rainy Season
Month
Original NDVI
Oct
No
Dec
Jan Feb
Dry Season
Mar Apr May Jun Jul Aug
Rainy Season
Sep
Oct
No
Dec
Dry
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000
Jan
Feb Mar Apr May
Dry Season
Jun
Jul Aug
Rainy Season
Sep
Oct No
v
Dec Jan Feb Mar
Dry Season
Apr May Jun
Jul
Aug Sep
Rainy Season
Month
Oct
No
v
Dec
Jan Feb
Dry Season
Mar Apr May Jun Jul Aug
Rainy Season
after Min Max Filter (WIN=3)
4
Sep
Oct
No
v
Dec
Dry
Background: LMF (cont.)
_____________________________________________________________________________________________________________________________________

Simulate the fluctuation by limited number of
Harmonic Waves.
 Fitting Model

 2 kl t 
 2 kl t  
ft  c0  c1t   c2l sin 

c
cos
 2l 1
 M 
 M 


l 1 
N

To avoid overfitting & Maximize prediction stability
 AIC
( Akaike Information Criterion )
AIC = D{log(2)+1}+2(j+1)
5
Statement of the Problems
________________________________________________________________________________________________________________________________________________________


LMF line is lower than maximum original NDVI (see
)
The number of peaks are different from real pattern
(sugarcane has 1 peak per cycle) (see
) Too large W : Revisable value
will exceed the real NDVI
Sugarcane: Comparison of Original NDVI and after LMF (WIN=3) from 2005 to 2007
NDVI
Too small W : Cannot remove noise
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000
Jan
Feb Mar Apr May
Dry Season
Jun
Jul
Aug Sep Oct No
v
Rainy Season
Dec Jan Feb Mar
Apr May Jun
Dry Season
Original NDVI


Month
Jul
Aug Sep
Oct
Rainy Season
No
v
Dec Jan Feb Mar Apr May Jun Jul Aug
Dry Season
Rainy Season
Sep
Oct
No
v
Dec
Dry
after LMF (WIN=3)
Data NDVI MODIS 8-days images of Suphanburi (Spatial resolution 250 m.)
Year 2005, 2006, 2007; 46 images/yr ( Total = 138)
Sugarcane ( Number of function = 6, Window size = 3 )
6
Objective
_____________________________________________________________________________________________________________________________________

To improve LMF algorithm
 To
adjust Min Max filter size according to the
weather condition.
 Use multiple previous years information for
repeating pattern in case similar pattern of same
type of crop.
(Best Max Value Method for revision data)
■
To improve a set of software
7
Methodology
______________________________________________________________________________________________________________________________________

Study Area : Suphanburi Province

Data




Modis surface reflectance products 8-Day L3 (250m.)(year 2005-2007)
 TERRA (MOD 09Q1), AQUA (MYD 09Q1) (46 images / year)
 Suphanburi occupies in 2 land tiles (Total = 552 images)
Ground Data (year 2003)
GIS land use map (year 2000)
Average monthly rainfall historical statistical data
8
Methodology: Overview
______________________________________________________________________________________________________________________________________
MODIS surface reflectance products
8-days(TERRA,AQUA)
1 Min Max Filter
Adjust Min Max window filter size
depend on
weather condition
Average Monthly
Historical Data
Rainfall
Cloud-days
Different pattern of
crop in each year
Similar pattern of
same type of crop
in each year
Improvement
NIR surface
reflectance
(TERRA)
Improvement :
Utilization of
2
Best Max Value
Method
Multi-year
Information
Harmonic
Curve Fitting
Model
3
4
Min AIC
Outcome
(Expected LMF)
9
1. Min Max Filter
_____________________________________________________________________________________________________________________________________________________
Adjust window size depending on weather condition
Concept 1: Depends on Average Monthly Rainfall Historical
Statistical Data
– Global : WMO, FAO, LocClim
– Local : Meteorological Department
Thai
Meteorological
Department
Average Monthly Rainfall from 1961-1990
300
Rainfall (mm.)
250
W=6
More Rainfall : Bigger Window Size
200
150
W=5
100
50
W=3
0
Jan
Feb
Mar
Dry Season
Apr
May
Jun
Jul
Month
Aug
Sep
Rainy Season
Heavy Rainy season
Oct
Nov
Dec
Dry
13
10
Min Max Filtering (Cont.)
________________________________________________________________________________________________________________________________________
Experiment: Varies Window Size depends on average rainfall data
NDVI
Min Max Filter
(WIN=3)
Comparison of NDVI before and after Min Max Filter (WIN=3) from 2005 to 2007
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000
Jan
Feb Mar
Dry Season
Apr
May
Jun
Jul
Aug
Rainy Season
Sep
Oct No
v
Dec Jan Feb Mar
Dry Season
Apr May Jun
Jul
Aug Sep
Rainy Season
Oct
No
Dec
Jan Feb
Dry Season
Mar
Apr
May Jun
Month
Heavy Rainy season
Original NDVI
before Min Max Filter
Jul
Aug
Sep
Oct
Rainy Season
No
v
Dec
No
v
Dec
Dry
after Min Max Filter
NDVI
Comparison of NDVI before and after Min Max Filter (WIN=5) from 2005 to 2007
Min Max Filter
(WIN=5)
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000
Jan
Feb Mar Apr
Dry Season
Original NDVI
May
Jun
Jul
Aug
Rainy Season
Sep
Oct No
v
Heavy Rainy season
Dec Jan Feb Mar
Dry Season
Apr May Jun
Month
before Min Max Filter
Jul
Aug Sep
Rainy Season
Oct
No
after Min Max Filter
Dec
Jan Feb
Dry Season
Mar
Apr
May Jun
Jul
Aug
Rainy Season
Sep
Oct
11
Dry
Min Max Filtering (cont.)
________________________________________________________________________________________________________________________________________
Result: Min Max Filter Graph
□
Varies Window Size depends on average rainfall data
Comparison of NDVI before and after Min Max Filter (WIN=3,5,6,5,3) from 2005 to 2007
NDVI
Min Max Filter
2000 (WIN=3,5,6,5,3
1900
)
1800
1700
1600
1500
1400
1300
1200
1100
1000
Jan
Feb Mar
Dry Season
Apr
May
Jun
Jul
Aug
Rainy Season
Sep
Oct No
v
Dec Jan Feb Mar
Dry Season
Apr May Jun
Jul
Aug Sep
Rainy Season
Oct
No
Month
Original NDVI
Heavy Rainy season
before Min Max Filter
Dry season
W=3
Rainy season
W=5
Heavy Rainy season
W=6
Dec
Jan Feb
Dry Season
Mar
Apr
May Jun
Jul
Aug
Sep
Oct
Rainy Season
No
v
after Min Max Filter
Mean Total Rainfall : F mean
0 ≤ F mean ≤ 100  w = 3
100 < F mean ≤ 200  w = 5
200 < F mean
 w=6
12
Dec
Dry
1. Min Max Filter (cont.)
____________________________________________________________________________________________________________________________________________
Concept 2 : Depends on continuous cloud-days
Define cloud-day: Cloud Mask Method
NIR surface reflectance (TERRA)
Cloud Mask
TERRA 2007 Band 43 (Dec.)
RGB=b2,b1,b2

NIR Reflectance Threshold = 3500
Cloud Mask
13
1. Min Max Filter (cont.)
_____________________________________________________________________________________________________________________________________________________
Concept 2 : Depends on continuous cloud-days (cont.)
N left, N right = The number of cont. cloud-days at the left or right side of considering point
Wleft , Wright = Window filter size at left, right side of considering point
Wleft = Nleft + 4
NDVI after filter = Min [ Max left , Max right ]
Wright = Nright + 4
Comparison of Original NDVI and after Min Max Filter (WIN=3 & WIN depends on cont. cloud-days) in 2006
Cloud-day =
Min Max Filter (WIN depends on cont. cloud-days)
2000
1900
1800
NDVI
1700
1600
1500
1400
1300
1200
Original NDVI
1100
Min Max Filter (WIN=3)
1000
1
9
Jan
17
25
33
41
Feb
49
57
65
73
Mar
81
89
97 105 113 121 129 137 145 153 161 169 177 185 193 201 209 217 225 233 241 249 257 265 273 281 289 297 305 313 321 329 337 345 353 361
Apr
May
Jun
DOY
Dry Season
Jul
Aug
Sep
Oct
Nov
Rainy Season
Dec
Dry
M onth
Min Max Filter (W depends on cloud-days) remove noise > W = 3
14
2. Best Max Value Method
_____________________________________________________________________________________________________________________________________________________
Stable Agriculture field  Repeat Same Pattern




Case : Similar pattern in each year
Choose maximum NDVI value within analysis
period
 Maximum value does not have cloud
More chance to capture good data
Justify similar or not similar pattern in each
year: Modeling Method (r 2)
15
2. Best Max Value Method (cont.)
_____________________________________________________________________________________________________________________________________________________
Comparison of NDVI after Min Max Filter and NDVI after adjust trend and
intersection point of NDVI axis of each year back
After adjust trend and intersection point of NDVI axis
back
NDVI Pattern
2000
A'
1900
B'
500
1800
Remove noise
using 3 yrs pattern
1600
1500
A
1400
B
1300
1200
1st
2nd
Year
C
400
300
200
NDVI
NDVI
1700
Min Max Filter has
3rd
trend

pattern
0
-100
-200
-300
-400
Year
1
Adjust graph each year to have no trend and intersection point of
NDVI axis


100
-500
NDVI after Min Max Filter
NDVI after adjust trend and intersection point of NDVI axis of each year back

pattern
C'
Calculate Slope (c1) , Remove trend ( c1t ) and intersection point of NDVI axis
(c0) by Linear Regression

NDVI no trend at time axis = NDVI after filter – c1t – c0
Choose maximum value within analysis period
pattern
Adjust graph each year to have trend and intersection point of NDVI
axis back

NDVI after adjust
trend and intersection point of NDVI axis back=
patternc0  c1t
16
Improved LMF Result
________________________________________________________________________________________________________________________________________
NDVI
Comparison of NDVI before and after LMF (WIN=3 & WIN=3,5,6,5,3,BestMax) from 2005 to 2007
LMF (WIN=3)
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000
LMF (WIN=3,5,6,5,3,Best Max)
Original NDVI
Jan
Feb Mar Apr May
Dry Season
Jun
Jul
Aug Sep Oct No
v
Rainy Season
Original NDVI
Dec Jan Feb Mar
Dry Season
Apr May Jun
Month
Jul
Aug Sep
Oct
Rainy Season
after LMF (WIN=3)
No
v
Dec
Jan Feb
Dry Season
Mar Apr May Jun Jul Aug
Sep
Oct
Nov Dec
Rainy Season
Dry
after LMF (WIN=3,5,6,5,3, BestMax)
NDVI
Comparison of Original NDVI and after LMF (WIN=3 & WIN depends on cont. cloud-days, Best Max) 2005-2007
LMF (WIN=3)
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000
LMF (WIN depends on cont. cloud-days, Best
Max)
Original NDVI
Jan
Feb
Mar
Dry Season
Apr
May
Jun
Jul
Aug
Rainy Season
Sep
Oct No
v
Original NDVI
Dec Jan Feb Mar
Dry Season
Apr May Jun
Month
after LMF (WIN=3)
Jul
Aug Sep
Rainy Season
Oct
No
v
Dec
Jan
Feb
Mar
Apr
May Jun
Dry Season
Jul
Aug
Rainy Season
Sep
Oct
No
v
Dec
Dry
after LMF (WIN depends on cont. cloud-days, Best Max)
Both Improved LMF are lifted higher, and approach original NDVI Top Points
17
Accuracy Assessment
_____________________________________________________________________________________________________________________________________
NDVI
Comparison of Original NDVI and Original NDVI Top Point from 2005 to 2007
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000
Jan
Feb Mar Apr May
Dry Season
Jun
Jul
Aug Sep Oct No
v
Rainy Season
Dec Jan Feb Mar
Dry Season
Apr May Jun
Month
Original NDVI
n
 y
1i
rmsd 
 y 2i 
i 1
n
2
Jul
Aug Sep
Rainy Season
Oct
No
v
Dec
Jan Feb
Dry Season
Rainy Season
Sep
Oct
No
v
Dec
Dry
Original NDVI (Top Point)
y1i = original top point NDVI
y2i = after LMF NDVI
n = number of original NDVI top point
Improved LMF (WIN depends on Rainfall data)*:
 Improved LMF (WIN depends on Cont. cloud-days)*:
 Prior LMF (One WIN = 3):
(* Including Best Max)

Mar Apr May Jun Jul Aug
rmsd = 36.88
rmsd = 37.04
rmsd = 44.92
18
Accuracy Assessment (cont.)
_____________________________________________________________________________________________________________________________________

Plot comparison NDVI before and after LMF varies
58 pixels classify by 12 types land use
Rainfall
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000
Comparison of Original NDVI and after LMF (WIN=3 & WIN=3,5,6,5,3, Best Max) from 2005 to 2007
NDVI
NDVI
Comparison of Original NDVI and after LMF (WIN=3 & WIN=3,5,6,5,3, Best Max) from 2005 to 2007
Jan Feb Mar Apr May Jun Jul Aug Sep Oct No Dec Jan Feb Mar
v
Dry Season
Rainy Season
Dry Season
Original NDVI
Apr May Jun
Month
Jul Aug Sep Oct No Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct
v
Rainy Season
Dry Season
Rainy Season
after LMF ( WIN = 3,5,6,5,3, Best Max )
Sugarcane
after LMF (WIN=3)
No Dec
v
Dry
2000
1900
1800
1700
1600
1500
1400
1300
1200
1100
1000
Jan Feb Mar Apr May Jun Jul Aug Sep Oct No Dec Jan Feb Mar
v
Dry Season
Rainy Season
Dry Season
Original NDVI
Apr May Jun
Month
Jul Aug Sep Oct No Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct
v
Rainy Season
Dry Season
Rainy Season
after LMF ( WIN = 3,5,6,5,3, Best Max )
after LMF (WIN=3)
Cassava
19
No Dec
v
Dry
Conclusion & Recommendation
W=3
2




1
3
Rainfall
– Global data acquisition
Cont. cloud-days – More complicate to identify cloud-day
– Uncertain obtained cloud-day
One window size – Cannot remove noise in longer
rainfall period
Recommendation – Improvement cloud-day identification
using more suitable method
20
21