Mapping Ash Tree Colonization in an Agricultural Moutain Landscape_ Investigation the Potential of Hyperspectral Imagery.pdf
1. Mapping Ash Tree Colonization in an
Agricultural Moutain Landscape: Investigation
the Potential of Hyperspectral Imagery
D. Sheeren1,2 , M. Fauvel1,2 , S. Ladet2 , A. Jacquin2 , G. Bertoni1,2 and
A. Gibon2
1
INP Toulouse - University of Toulouse
2
UMR 1201 DYNAFOR, INRA/INP-ENSAT/INP-EI Purpan
2011 IEEE International Geoscience and Remote Sensing Symposium
24-29 July, Vancouver, Canada
4. Context of the work 1/3
Scientific context:
Landscape ecology:
What is the impact of agricultural activities on the landscape?
What is the impact of landscape heterogeneity on the biodiversity?
Global change:
How evolve landscape and biodiversity?
What are the factors of evolution?
Predict and anticipate the responses of ecosystems to landscape changes.
5. Context of the work 2/3
What are the causes and the consequences of ash three colonization in the
Pyrenees mountain?
Villelongue village, 65260-France
1950 2000
Up to now:
Ecological process understood
Multi-agents model build
× Accurate ash-thematic map
6. Context of the work 3/3
Current method: Aerial images + visual inspection + field survey
Small geographical area
Time consuming
Cost !
Multispectral satellite images are not enough spatially and spectrally
accurate for ash detection
tree/not tree ok, but it is not possible to go to the species
Good multitemporal data are very difficult to obtain
Hypothesis: With hyperspectral images, it would be possible to
differentiate between ash tree and other species of tree
7. Context of the work 3/3
Current method: Aerial images + visual inspection + field survey
Small geographical area
Time consuming
Cost !
Multispectral satellite images are not enough spatially and spectrally
accurate for ash detection
tree/not tree ok, but it is not possible to go to the species
Good multitemporal data are very difficult to obtain
Hypothesis: With hyperspectral images, it would be possible to
differentiate between ash tree and other species of tree
Madonna Project!
8. Madonna project: Objectives and data
Objectives:
• Mapping of the ash tree distribution
• 2D et 3D information
• Estimation of structural and biophysical parameters (tree density and height,
foliar chlorophyll . . . )
Data (summer 2010).
• Very high spatial resolution hyperspectral images
• LiDar data
• Field data (ash trees and other dominant species, foliar analysis . . . )
9. Area covered by the mission
Village of Villelongue, France (00◦ 03’W and 42◦ 57’N).
Medium altitudinal range (450-1800m)
I
14. Very high spatial resolution hyperspectral images 1/2
T
E
HySpex sensor:
Spectral resolution : 1.5 nm, 400-1000 nm, 160 bands
Spatial resolution: 50 cm
15. Very high spatial resolution hyperspectral images 2/2
Pattern recognition approach:
Problem of the spectral dimensionality: statistical methods fail
Use of non-linear SVM (Gaussian kernel)
But the optimization of the hyperparameter is too demanding in terms of
time processing when using the conventional cross-validation strategy
(about 3 To of data to process).
A fast and accurate method for optimizing the hyperparameter is needed
for an operational system
SVM + Kernel alignment
16. Very high spatial resolution hyperspectral images 2/2
Pattern recognition approach:
Problem of the spectral dimensionality: statistical methods fail
Use of non-linear SVM (Gaussian kernel)
But the optimization of the hyperparameter is too demanding in terms of
time processing when using the conventional cross-validation strategy
(about 3 To of data to process).
A fast and accurate method for optimizing the hyperparameter is needed
for an operational system
SVM + Kernel alignment
17. Support Vector Machine
H (α, b) : {x|f (x) = 0}
yi = 1
w
yi = −1
Supervised method: S = (x1 , y1 ), . . . , (xn , yn ) ∈ Rd × {−1; 1}
n
Separating function: f (z) = sgn αi k(z, xi ) + b
i=1
Solve QP problem: n n
1
max g(α) = αi − αi αj yi yj k(xi , xj )
α
i=1
2 i,j=1
n
constraint to 0 ≤ αi ≤ C et i=1 αi yi = 0
18. Kernel function
Kernel function k: similarity measure between two spectra xi et xj
2
xi − xj
Gaussian kernel : kg (xi , xj ) = exp −
2σ 2
σ 2 is an hyperparameter that controls how two spectra are considered as
similar or not.
Ideally, σ 2 must be tuned such as:
kg (xi , xj ) ≈ 1 If yi = yj
kg (xi , xj ) ≈ 0 Else
Ideal kernel matrix:
1 δ y1 y2 . . . δy1 yn
δ y2 y1 1 . . . δy2 yn
I
K = .
. . .. .
. .
. . .
.
δyn y1 δyn y2 ... 1
19. Kernel alignment
Alignment A: compute the similitude (angle) between the ideal matrix and
the kernel matrix with parameter σ 2
KI, K F
A(σ) =
KI F K F
σ is selected such A(σ) is maximal
Contrary to cross-validation, there is no need to solve the QP problem
0.5
0.45
0.4
0.35
0.3
0.25
0 20 40 60
21. Protocol
Ground thruth:
12 tree species (Ash tree, Chestnut tree, Lime tree, Hazel tree . . . ).
Classification of the tree species: Are ash trees identifiable?
Classification of the image: Are the results spatially consistent?
22. Results
Quantitative analysis: Ash tree separability
GMM SVM SVM-lin
OA 72% 94% 89%
Kappa 0.65 0.92 0.89
Ash tree
User accuracy 84.0% 89.9% 83.1%
Producer accuracy 53.6% 89.9% 88.8%
Qualitative analysis:
24. Conclusions and perspectives
Conclusions:
Accurate mapping of ash tree is possible with hyperspectral images
Framework: SVM + kernel alignment
Perspectives:
Spatial regularization
Can biophysical parameters be estimated?
25. Mapping Ash Tree Colonization in an
Agricultural Moutain Landscape: Investigation
the Potential of Hyperspectral Imagery
D. Sheeren1,2 , M. Fauvel1,2 , S. Ladet2 , A. Jacquin2 , G. Bertoni1,2 and
A. Gibon2
1
INP Toulouse - University of Toulouse
2
UMR 1201 DYNAFOR, INRA/INP-ENSAT/INP-EI Purpan
2011 IEEE International Geoscience and Remote Sensing Symposium
24-29 July, Vancouver, Canada