Capturing ecological properties of fragmented landscapes
Novel image processing based methods
v.lefebvre@imperial.ac.uk
Véronique Lefebvre, Marion Pfeifer, Andrew Bradley and Robert Ewers
What PatchStats can do for landscape ecology
1
Characterisation of fragmented landscapes
Delineating fragments by Watershed segmentation
Fragmentation affects biodiversity. To understand and predict
changes in biodiversity with fragmentation we need to assess
the spatial configuration of landscapes.
2
The delineation method estimates if a fragment should be sectioned, and if so which pixels should be separated.
Edge for species 1
Edge for species 2
Watershed
line
This poster presents novel computational solutions to:
Fragment
The Watershed method enables segmenting landscape maps into ecologically meaningful fragments.
The Compactness and Smoothness descriptors assess distinct properties of fragments shape.
These “PatchStats” techniques can be widely applied in landscape ecology and will be used to model
biodiversity responses to forest habitat.
Catchment basins
Delineation
1 delineate habitat fragments
a
2 characterise their shape
Shape classification
a
Area
Cut for species 2
Method: Watershed segmentation
Task: Segment shape
+
7
6
1
3
Shape Factor
Algorithm’s steps
Distance
to forest
edge
1
a
9
2
8
5
Compactness
a
Distance to forest edge map
3D view
-
Flatten local maxima < r
H-maxima transform
Cut everything > T
Thresholding
4
10
-
+
-
+
1
Distance to edge
Example of data:
Binary map of forest in the Comoros Islands
Fractal
Dimension
50
100
150
200
250
300
350
400
450
500
Map segmented into fragments by the Watershed
method.
Smoothness
2
Classifications by Shape Factor and to a lesser extent by Fractal Dimension reflect fragments area.
Fractal Dimension assesses both the dimension of the contour and the compactness.
Compactness and Smoothness are independent and characterise distinct shape properties.
Fragments Shape
Compactness
Invert
Watershed
segmentation
Fragments = Basins
2
New descriptors
Variations of common and new descriptors with area
Perimeter
Smoothness
Parameters meaning and tuning to match ecological background
Thin parts in fragments are cut if:
- The fragment contains core forest
i.e. if max(distance to edge in fragment) > r
- The thinner part is thin enough
i.e. if min(distance to edge in thin part) < T
Different
types of
habitat
no core
forest no cut
T and r can be adjusted to map resolution.
T gives a direct control on the maximum forest section
that the algorithm can cut.
Too thick to
be cut
T
T and r combined reflect the depth of edge influence.
They can be tuned to obtain a segmentation that matches
any prior ecological knowledge of the area.
5
The perimeter is a noisy variable in pixel based maps.
5
3
2
0
1
10
Computing the perimeter by joining up the
mid-points of border pixel sides minimises the
bias when no interpolation is performed.
0
10
Smoothness
0.9
0.8
It leads to the octagon
having the smallest
perimeter to area ratio.
0.5
0.7
0.6
0.5
10
0
10
Problems with common shape descriptors
Fractal Dimension
2
same area
same perimeter
How plane-filling is the contour of the shape?
Smoothness
Compactness
How wriggly is the contour line of a shape?
How many indents in its contour?
How packed is the shape?
same descriptor
circle, rectangle
4
Perimeter
same simple contour
descriptors
depend on area
Different Fractal dimension
Fractal dimension also reflects compactness
Number
of indents
2
1
0
-1
Fragment Area
Fragment Area
0
1
2
3
Fragment area
4
1
2 zero crossings
dx
Fractal Dimension
Problems for
pixel-based
maps
3
Complexity
Compactness
Fragment Shape
complex shape
indents
Proportion of
smooth perimeter
0
-1
0
5
10
15
20
25
30
35
40
45
50
1
8 zero crossings
dy
How compact is a shape compared to a
circle of same area?
0
Fragment area
coordinates derivative
Polygon Shape Factor
0
Compactness
1
0
2
Fractal Dimension
4
Counting the sides of border pixels can overestimate the perimeter
by a factor √2, introducing a bias proportional to area.
It also implies that the shape with the smallest perimeter to area ratio
is the square.
r
Shape Factor
10
0
-1
0
5
10
15
20
25
30
35
5
Perimeter length
40
45
50