Detection and mapping of forest changes : application to the ...
Detection and mapping of forest changes : application to the ...
Detection and mapping of forest changes : application to the ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
BEGUET Benoît (Université de Bordeaux <strong>and</strong> INRA)<br />
benoit.beguet@egid.u-bordeaux3.fr<br />
BOUKIR Samia (ENSEGID)<br />
CHEHATA Nesrine (ENSEGID)<br />
GUYON Dominique (INRA)<br />
<strong>Detection</strong> <strong>and</strong> <strong>mapping</strong> <strong>of</strong> <strong>forest</strong> <strong>changes</strong> :<br />
<strong>application</strong> <strong>to</strong> <strong>the</strong> maritime pine st<strong>and</strong>s in SW France<br />
Pleiades Days<br />
Toulouse, 17-18 january 2012<br />
1
Context<br />
<br />
The ORFEO Accompaniment Program :<br />
<strong>to</strong> prepare, accompany <strong>and</strong> promote <strong>the</strong> use <strong>and</strong> <strong>the</strong> exploitation <strong>of</strong> <strong>the</strong> images<br />
derived from Pleiades sensors.<br />
Group GT6-Forêt<br />
<br />
B.Beguet <strong>the</strong>sis (oct 2010 – oct 2013):<br />
« <strong>Detection</strong>, characterization <strong>and</strong> moni<strong>to</strong>ring <strong>of</strong> <strong>forest</strong> <strong>changes</strong> from Very High<br />
Resolution satellite imagery »<br />
<br />
ENSEGID / INRA : collaboration since 2009<br />
2
Introduction : general overview<br />
...Very High Resolution satellite imagery...<br />
<br />
Which <strong>the</strong>matic innovation could we expect ?<br />
<br />
Which methodological innovation is needed ?<br />
<br />
Our priority : Forest <strong>changes</strong><br />
<br />
Strong damages : windfall, massive dieback<br />
<br />
Anthropogenic activities : de<strong>forest</strong>ation, re<strong>forest</strong>ation ...<br />
3
Introduction : HR vs VHR<br />
Spatial resolution :<br />
<br />
defines <strong>the</strong> dimension <strong>of</strong> <strong>the</strong> observed (or perceptible) objects<br />
<br />
HR (Spot5, Formosat) : usefull for many <strong>application</strong>s but :<br />
pixel size > tree crown size<br />
→ resolution is <strong>to</strong> low (5-10m) <strong>to</strong> describe <strong>forest</strong> st<strong>and</strong>s complexity <strong>and</strong> diversity<br />
<br />
VHR (QuickBird, World-View, Pleiades), relation between :<br />
{pixel size (0.5-2.5m)} <strong>and</strong> {tree crown size (0-8m)}<br />
→ <strong>forest</strong> is no longer a « green carpet » : TEXTURE (tree's spatial organization)<br />
How <strong>to</strong>? → Optimize existing methods by an approriate use <strong>of</strong> VHR<br />
imagery textural information at <strong>the</strong> st<strong>and</strong> level (object-based)<br />
4
Outlines<br />
1. Application <strong>of</strong> interest :<br />
An au<strong>to</strong>matic frame for wind-fall damage binary <strong>mapping</strong> in <strong>forest</strong>s<br />
2. On <strong>the</strong> use <strong>of</strong> VHR :<br />
Retrieving <strong>forest</strong> structure variables from VHR imagery<br />
5
1. Application <strong>of</strong> interest<br />
Change detection : <strong>application</strong> <strong>to</strong> <strong>the</strong> windfall damage, Nezer <strong>forest</strong>, 2009<br />
OBJECT-BASED FOREST CHANGE DETECTION USING HIGH RESOLUTION<br />
SATELLITE IMAGES, C.Orny, N.Chehata, S. Boukir, D.Guyon (PIA 2011)<br />
<br />
<br />
Objective : provide an au<strong>to</strong>matic method for damage binary <strong>mapping</strong><br />
Data : Formosat2 (8m spatial resolution, before <strong>and</strong> after s<strong>to</strong>rm)<br />
<br />
Method :<br />
→ Object based (st<strong>and</strong> level )<br />
→ Au<strong>to</strong>matic feature selection using test frames<br />
→ Mean Shift segmentation & classification<br />
→ Change detection by joint classification<br />
→ Novel spatio temporal descrip<strong>to</strong>r: fragmentation rate<br />
6
1. Application <strong>of</strong> interest<br />
Results :<br />
Global<br />
Accuracy<br />
87.2 %<br />
Before s<strong>to</strong>rm<br />
NO CHANGES TYPOLOGY<br />
After s<strong>to</strong>rm<br />
Au<strong>to</strong>matic classification<br />
intact<br />
damage<br />
Reference data<br />
intact<br />
damage<br />
7
2. VHR, texture <strong>and</strong> <strong>forest</strong> structure<br />
Spot5 MS : 10m QuickBird MS : 2.4m QuickBird Pan : 0.6m<br />
If it could be easy <strong>to</strong> recognize a texture (with human eyes <strong>and</strong> brain), <strong>the</strong>re is no unique<br />
(formal) way <strong>to</strong> define it...<br />
<br />
Texture description <strong>the</strong>ory :<br />
→ First order statistics : mean, variance, higher order moments<br />
→ Second order statistics : Grey Level Cooccurence Matrix (GLCM) <strong>and</strong> <strong>the</strong><br />
derived Haralick's features (energy , entropy , correlation ... (14))<br />
→ Semi-variogramm : sill, range <strong>and</strong> nugget<br />
→ Frequential approach: TFourier , Gabor filters , Wavelets<br />
→ O<strong>the</strong>r : geometric approach (TPN) ; ma<strong>the</strong>matical morphology<br />
<br />
Parametrization :<br />
→ common ground: computation window size → scale → objects size<br />
8
2. VHR, texture <strong>and</strong> <strong>forest</strong> structure<br />
GLCM <strong>the</strong>ory, a quick look :<br />
For each pixel :<br />
135°<br />
90°<br />
yd<br />
45°<br />
Count <strong>the</strong> number <strong>of</strong> each grey level pairs :<br />
- in a window with fixed size (w)<br />
- with a fixed displacement <strong>and</strong> orientation (<strong>of</strong>fset defined by (xd;yd))<br />
- for a given level <strong>of</strong> quantification<br />
Exemple : w = 1 ; xd = 1 = yd (orientation=45°)<br />
2<br />
1<br />
0 1 2<br />
xd<br />
0°<br />
image<br />
GLCM for <strong>the</strong> « red » pixel<br />
1 1<br />
3<br />
1<br />
2<br />
ng 1<br />
2<br />
3 4<br />
1<br />
3<br />
4<br />
1<br />
4<br />
1<br />
2 0 0 1<br />
2<br />
4<br />
1<br />
3 3<br />
2<br />
0<br />
0 0 0<br />
Haralick's features<br />
3<br />
1<br />
3<br />
4<br />
3 3<br />
0 0<br />
1<br />
0<br />
2<br />
2<br />
1<br />
1<br />
2<br />
4<br />
1<br />
0 0<br />
1<br />
PROBLEM : determination <strong>of</strong> <strong>the</strong> optimal set <strong>of</strong> parameters for a given <strong>application</strong><br />
SOLUTION : test a wide range <strong>of</strong> parameter values ( au<strong>to</strong>matization )<br />
9
2. VHR, texture <strong>and</strong> <strong>forest</strong> structure<br />
Retrieving <strong>forest</strong> structure variables from QuickBird imagery using an au<strong>to</strong>matic<br />
method :<br />
<br />
Ground measured <strong>forest</strong> variables :<br />
tree height (Ht), crown diameter (Cd), st<strong>and</strong> density (Nah), diameter at breast height<br />
(Dbh), basal area (Ba), tree spacing (Sp) on 12 sample plots.<br />
<br />
St<strong>and</strong> age is known on <strong>the</strong> whole site → used <strong>to</strong> method validation<br />
QuickBird image 06 oct 2003 : Pan (0.6m), Ms (2.4m) (Pleiades like)<br />
<br />
Method :<br />
<br />
<br />
Linear regressions : find <strong>the</strong> strongest relation between each<br />
<strong>forest</strong> variable <strong>and</strong> each image features<br />
Test frames : allows a calculation <strong>of</strong> thous<strong>and</strong>s <strong>of</strong> features<br />
<br />
Towards a multi-scale approach :<br />
→ combine Pan <strong>and</strong> Ms features ; combine different parametrizations<br />
→ multi-linear regressions problem : multicollinearity<br />
→ solution : clever au<strong>to</strong>matic subset selection<br />
LARS (Least Angular RegressionS) stepwise procedure<br />
10
2. VHR, texture <strong>and</strong> <strong>forest</strong> structure<br />
<br />
Test frames dimensions : 124*124m (200*200pix Pan, 50*50pix MS) :<br />
→ Representative dimension <strong>of</strong> <strong>the</strong> within-st<strong>and</strong> variability<br />
→ Correspondance with ground measurement plots<br />
St<strong>and</strong> ages<br />
13 15 22 23 26 26<br />
11<br />
30 32 43 48 48 51
2. VHR, texture <strong>and</strong> <strong>forest</strong> structure<br />
Results : single feature solutions<br />
- Strong linear relationships between <strong>forest</strong><br />
structure variables <strong>and</strong> optimized features<br />
- Pan <strong>and</strong> Ms provide equivalent performances<br />
- The few number <strong>of</strong> observations doesn' t allow<br />
a parameter interpretation<br />
12
2. VHR, texture <strong>and</strong> <strong>forest</strong> structure<br />
Results : multiple features solutions<br />
- 3 variables<br />
- no multicollinearity (VIF ~ 1)<br />
- all scores are improved (all<br />
multiple R² > 0.98 )<br />
- combination <strong>of</strong> Pan <strong>and</strong> Ms<br />
features<br />
- combination <strong>of</strong> different<br />
parametrizations (window size)<br />
Multi-scale predic<strong>to</strong>r<br />
13
2. VHR, texture <strong>and</strong> <strong>forest</strong> structure<br />
Results : validation with st<strong>and</strong> ages (184 st<strong>and</strong>s)<br />
Single solutions<br />
Panchromatic is more relevant → resolution importance<br />
Pseudo-optimal variables subset<br />
RMSE = 5.6 years<br />
RMSE% = 11.5%<br />
RMSE = 4.3 ans<br />
RMSE% = 8%<br />
Noticeable performance <strong>of</strong> features combinations<br />
14
2. VHR, texture <strong>and</strong> <strong>forest</strong> structure<br />
...Waiting for <strong>the</strong> classification step...<br />
affectation <strong>of</strong> <strong>the</strong> predicted age on <strong>the</strong> whole st<strong>and</strong>, 5 classes<br />
2. VHR, texture <strong>and</strong> <strong>forest</strong> structure<br />
Conclusions :<br />
<br />
<br />
<br />
<br />
The texture analysis allows <strong>to</strong> retrieve <strong>the</strong> <strong>forest</strong> structure. Results consistent<br />
with literature.<br />
GLMC's features needs <strong>to</strong> be optimized → au<strong>to</strong>matization thanks <strong>to</strong> test frames<br />
The au<strong>to</strong>matic features selection method find optimal features combinations <strong>to</strong><br />
retrieve <strong>forest</strong> structure variables. This produces a multi-scale descrip<strong>to</strong>r.<br />
The proposed method is generic, it could be applied for o<strong>the</strong>r <strong>application</strong>s<br />
(estimation <strong>of</strong> <strong>forest</strong> structure from VHR images texture)<br />
Perspectives :<br />
<br />
Validation : o<strong>the</strong>r sites, o<strong>the</strong>r sensors, o<strong>the</strong>r seasons ....<br />
<br />
Classification <strong>of</strong> <strong>forest</strong> st<strong>and</strong>s structure<br />
16
Perspectives<br />
<br />
Robustness, reproducibility <strong>and</strong> performance need <strong>to</strong> be tested with<br />
Pleiades images<br />
<br />
Fine change detection needs temporal series (few VHR data are available)<br />
waiting for Pleiades series<br />
<br />
Finest informations could also be exploited for <strong>forest</strong> structure<br />
characterization (not at st<strong>and</strong> level but at tree level)<br />
e.g. : Olivier Regniers <strong>the</strong>sis, Geometric <strong>and</strong> textural descrip<strong>to</strong>rs for Very High<br />
Resolution remote sensing (Bordeaux Sciences Agro / IMS / Université de<br />
Bordeaux)<br />
17
Beautiful <strong>and</strong> powerful Pleiades images are expected !<br />
Tagon<br />
Marcheprime<br />
Nezer<br />
KALIDEOS<br />
18