Detection and mapping of forest changes : application to the ...

BEGUET Benoît (Université de Bordeaux and INRA)
benoit.beguet@egid.u-bordeaux3.fr
BOUKIR Samia (ENSEGID)
CHEHATA Nesrine (ENSEGID)
GUYON Dominique (INRA)
Detection and mapping of forest changes :
application to the maritime pine stands in SW France
Pleiades Days
Toulouse, 17-18 january 2012
1

Context
The ORFEO Accompaniment Program :
to prepare, accompany and promote the use and the exploitation of the images
derived from Pleiades sensors.
Group GT6-Forêt
B.Beguet thesis (oct 2010 – oct 2013):
« Detection, characterization and monitoring of forest changes from Very High
Resolution satellite imagery »
ENSEGID / INRA : collaboration since 2009
2

Introduction : general overview
...Very High Resolution satellite imagery...
Which thematic innovation could we expect ?
Which methodological innovation is needed ?
Our priority : Forest changes
Strong damages : windfall, massive dieback
Anthropogenic activities : deforestation, reforestation ...
3

Introduction : HR vs VHR
Spatial resolution :
defines the dimension of the observed (or perceptible) objects
HR (Spot5, Formosat) : usefull for many applications but :
pixel size > tree crown size
→ resolution is to low (5-10m) to describe forest stands complexity and diversity
VHR (QuickBird, World-View, Pleiades), relation between :
{pixel size (0.5-2.5m)} and {tree crown size (0-8m)}
→ forest is no longer a « green carpet » : TEXTURE (tree's spatial organization)
How to? → Optimize existing methods by an approriate use of VHR
imagery textural information at the stand level (object-based)
4

Outlines
1. Application of interest :
An automatic frame for wind-fall damage binary mapping in forests
2. On the use of VHR :
Retrieving forest structure variables from VHR imagery
5

1. Application of interest
Change detection : application to the windfall damage, Nezer forest, 2009
OBJECT-BASED FOREST CHANGE DETECTION USING HIGH RESOLUTION
SATELLITE IMAGES, C.Orny, N.Chehata, S. Boukir, D.Guyon (PIA 2011)
Objective : provide an automatic method for damage binary mapping
Data : Formosat2 (8m spatial resolution, before and after storm)
Method :
→ Object based (stand level )
→ Automatic feature selection using test frames
→ Mean Shift segmentation & classification
→ Change detection by joint classification
→ Novel spatio temporal descriptor: fragmentation rate
6

1. Application of interest
Results :
Global
Accuracy
87.2 %
Before storm
NO CHANGES TYPOLOGY
After storm
Automatic classification
intact
damage
Reference data
intact
damage
7

2. VHR, texture and forest structure
Spot5 MS : 10m QuickBird MS : 2.4m QuickBird Pan : 0.6m
If it could be easy to recognize a texture (with human eyes and brain), there is no unique
(formal) way to define it...
Texture description theory :
→ First order statistics : mean, variance, higher order moments
→ Second order statistics : Grey Level Cooccurence Matrix (GLCM) and the
derived Haralick's features (energy , entropy , correlation ... (14))
→ Semi-variogramm : sill, range and nugget
→ Frequential approach: TFourier , Gabor filters , Wavelets
→ Other : geometric approach (TPN) ; mathematical morphology
Parametrization :
→ common ground: computation window size → scale → objects size
8

2. VHR, texture and forest structure
GLCM theory, a quick look :
For each pixel :
135°
90°
yd
45°
Count the number of each grey level pairs :
- in a window with fixed size (w)
- with a fixed displacement and orientation (offset defined by (xd;yd))
- for a given level of quantification
Exemple : w = 1 ; xd = 1 = yd (orientation=45°)
2
1
0 1 2
xd
0°
image
GLCM for the « red » pixel
1 1
3
1
2
ng 1
2
3 4
1
3
4
1
4
1
2 0 0 1
2
4
1
3 3
2
0
0 0 0
Haralick's features
3
1
3
4
3 3
0 0
1
0
2
2
1
1
2
4
1
0 0
1
PROBLEM : determination of the optimal set of parameters for a given application
SOLUTION : test a wide range of parameter values ( automatization )
9

2. VHR, texture and forest structure
Retrieving forest structure variables from QuickBird imagery using an automatic
method :
Ground measured forest variables :
tree height (Ht), crown diameter (Cd), stand density (Nah), diameter at breast height
(Dbh), basal area (Ba), tree spacing (Sp) on 12 sample plots.
Stand age is known on the whole site → used to method validation
QuickBird image 06 oct 2003 : Pan (0.6m), Ms (2.4m) (Pleiades like)
Method :
Linear regressions : find the strongest relation between each
forest variable and each image features
Test frames : allows a calculation of thousands of features
Towards a multi-scale approach :
→ combine Pan and Ms features ; combine different parametrizations
→ multi-linear regressions problem : multicollinearity
→ solution : clever automatic subset selection
LARS (Least Angular RegressionS) stepwise procedure
10

2. VHR, texture and forest structure
Test frames dimensions : 124*124m (200*200pix Pan, 50*50pix MS) :
→ Representative dimension of the within-stand variability
→ Correspondance with ground measurement plots
Stand ages
13 15 22 23 26 26
11
30 32 43 48 48 51

2. VHR, texture and forest structure
Results : single feature solutions
- Strong linear relationships between forest
structure variables and optimized features
- Pan and Ms provide equivalent performances
- The few number of observations doesn' t allow
a parameter interpretation
12

2. VHR, texture and forest structure
Results : multiple features solutions
- 3 variables
- no multicollinearity (VIF ~ 1)
- all scores are improved (all
multiple R² > 0.98 )
- combination of Pan and Ms
features
- combination of different
parametrizations (window size)
Multi-scale predictor
13

2. VHR, texture and forest structure
Results : validation with stand ages (184 stands)
Single solutions
Panchromatic is more relevant → resolution importance
Pseudo-optimal variables subset
RMSE = 5.6 years
RMSE% = 11.5%
RMSE = 4.3 ans
RMSE% = 8%
Noticeable performance of features combinations
14

2. VHR, texture and forest structure
...Waiting for the classification step...
affectation of the predicted age on the whole stand, 5 classes

2. VHR, texture and forest structure
Conclusions :
The texture analysis allows to retrieve the forest structure. Results consistent
with literature.
GLMC's features needs to be optimized → automatization thanks to test frames
The automatic features selection method find optimal features combinations to
retrieve forest structure variables. This produces a multi-scale descriptor.
The proposed method is generic, it could be applied for other applications
(estimation of forest structure from VHR images texture)
Perspectives :
Validation : other sites, other sensors, other seasons ....
Classification of forest stands structure
16

Perspectives
Robustness, reproducibility and performance need to be tested with
Pleiades images
Fine change detection needs temporal series (few VHR data are available)
waiting for Pleiades series
Finest informations could also be exploited for forest structure
characterization (not at stand level but at tree level)
e.g. : Olivier Regniers thesis, Geometric and textural descriptors for Very High
Resolution remote sensing (Bordeaux Sciences Agro / IMS / Université de
Bordeaux)
17

Beautiful and powerful Pleiades images are expected !
Tagon
Marcheprime
Nezer
KALIDEOS
18