pattern-based assessment of 2001/2006 land...
TRANSCRIPT
PAT TERN-BASED ASSESSMENT OF 2001/2006 L AND COVER CHANGE OVER THE ENTIRE U.S.Pawel Netzel and Tomasz F. Stepinski
Introduction
Change map
Methods
[email protected] [email protected]
Poster presented at the IGARSS 2014, July 13-18, 2014, Quebec, Canada
1,2 1 Space Informatics Lab (http://sil.uc.edu/) University of Cincinnati, USA
Dept. of Climatology and Atm. Protection, University of Wroclaw, Poland
1
2
2001
a
2006
bc
de
1 2 3 4
ab
cd
e1 2 3 4 scene a3 at 2006
scene a3 at 2001
0.0
0.2
0.4
0.55
0.0
0.2
0.4
0.55774 clumps
1141 clumps
land cover classes
land cover classes
clump-size classes
clump-size classesfrac
tion
of c
ells
NLC
D 2
001
NLC
D 2
006
land cover pattern clumping
frac
tion
of c
ells
A
B
C
D
E
<1.0 <0.99 <0.9 <0.8 <0.7 <0.6 <0.5 <0.4
similarity
change 2006 2001
change 2006 2001
change 2006 2001
developedhigh intensity
developedmed. intensity
developedlow intensity
developedopen space
cultivatedcrops woodywetlands
open water
deciduousforest
pasture/hay
evergreenforest
ice/snow
barren land
mixedforest
shrub/scrub
grassland
emergentwetlands
We present a post-classification change detection method geared toward assessing land cover
change on continental scale . Instead of tracking transitions of land cover classes on pixel-by-pixel basis the method measures the change in local patterns defined on 4.5 x 4.5 km square scenes (see illustration bellow). A pattern in a scene is represented by a 2D histogram of land cover classes and clump sizes and the level of change is measured as the dis-
similarity between motifs of scene patterns at two time points the Jensen-Shannon similarity measure. The methodology is applied to the National Land Cover Dataset (NLCD) to obtain a 2001-2006 change map of the conterminous U.S.
Fig.1 shows co-registered scenes at
two time points; land cover change is visible. Fig.2 shows (in random colors) clumps - contiguous groups of same-category pixels. Each pixel is assigned two variables - land cover category (one of 16 possible) and clump size category inherited from a clump to which it belongs (one of 14 possible). Clump size categories are categorized clump sizes. Fig.3 shows histograms of scenes pixels with re-spect to the two variables.
Dissimilarity between two histograms A and B is calculated using the Jensen-Shannon divergence (JSD), where H is en-tropy.
Fig.1 Fig.2 Fig.3
JSD(A,B) = H A+B2( ( 1
2 [H(A)+H(B)]-
Map of 2001-2006 pattern-based land cover change over the conterminous U.S. The map is 1045 x 1612 grid with each pixel represent-ing a value of JS similarity (1-JSD) between the same local scene in 2006 and 2001. The smaller the similarity the bigger the change. Most of the U.S. experienced little land cover pattern change between 2001 and 2006 (blue color), however, there are regional exceptions including the southeastern and Gulf regions, the Pacific Northwest region, and the state of Maine. There are also multiple local regions showing change in pattern. Five examples are indicated by letters A to E.
change 2006 2001A
The location of the 2002 Hyman forest fire in Colorado [15]. NLCD shows that this area has been covered by a well-consolidated forest in 2001, but the 2006 map shows a scar left by the fire. Pattern change map shows this location as a red spot.
B
C
change 2006 2001D
The location denoted by the letter Bcoincides with the Great Salt Lake in Utah. The change is due to the lake retreated from its 2001 levels to expose more “barren land” in 2006.
Locations denoted by the letter C corre-sponds to urban growth in the cities of Las Vegas, NV and Phoenix. Pattern change map clearly shows expansion of Las Vegas.
ELinear features on the change map correspond to rivers and their surroundings. In particular, letter E denotes the Rio Grande river in New Mexico. The portion of the Rio Grande river and the associated change in land cover be-tween 2001 and 2006 are shown. The changes can be ex-plained by fluctuating water levels.
Locations denoted by the letter D show from left to right) cities of Kansas City, St. Louis, Indianapolis, and Co-lumbus. These cities experi-enced some moderate level of urban development in their suburbia. Indianapolis is shown.
NLCDclasses
2006
2001
0.4 0.64 0.58 0.63 0.71 0.71 0.74 0.85 0.84 0.88 0.92 0.93 0.36 0.63 0.67 0.62 0.87 0.79 0.74 0.89 0.88 0.86 0.93 0.96
2006
2001
A B
11 12 21 22 23 24 31 41 42 43 52 71 81 82 90 950.0
0.1
0.2
0.3
0.4
0.5
0.6
11 12 21 22 23 24 31 41 42 43 52 71 81 82 90 950.0
0.1
0.2
0.3
0.4
0.5
0.6
0.51 0.89 0.48
42 43 52 710.0
0.1
0.2
0.3
0.4
0.5
0.6
42 43 52 710.0
0.1
0.2
0.3
0.4
0.5
0.6
0.10 0.79 0.74
A B
2001 2001
2006 2006
Comparison of different measures of change between scenesIn addition to JSS other methods of scene change are: JSS1 which uses only histograms of land cover classes and capture only changes in class composi-tion of the scene, and the fraction of unchanged pixels (r) which measures percentage of pixels in the scene that did not change land cover label. All measures have values between 0 and 1 but yield different values for the same pair of scenes. Figure to the right shows comparison of the values of these three measures in urban environment (Fig.A) and rural environment (Fig.B)
JSS JSS1 r JSS JSS1 r JSS JSS1 r JSS JSS1 r
Fig. A shows a scatter plot of r vs. JSS1 values for all scenes in the change map. The upper-right corner of the plot groups scenes that show very little change, whereas the lower-left corner of the plot groups tiles that show massive change. The tiles are cen-tered around the diagonal of the plot indicating that, on average, both measures have about the same sensitivity. However, there are outliers; P1 indicates one such outlier.Fig. B shows a scatter plot of JSS1 vs. JSS values for all scenes in the change map. The values of JSS are always smaller or equal to the values of JSS1 indicating that patterns are more susceptible to change than bulk compositions. P2 indicates one scene with high value of JSS1 but low value of JSS.
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
P2
B
0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.2
0.4
0.6
0.8
1.0
P1A
Similarity between bulk composition (JSS1)Fraction of unchanged pixels (r)
Sim
ilarit
y be
twee
n bu
lk c
ompo
sitio
n (J
SS1)
Sim
ilarit
y be
twee
n pa
tter
ns (J
SS)
Fig.A focuses on outlier P1 - the scene where JSS indicates very little change but about 50% of pixels changed labels. The change in this tile is due to the cyclic nature of forest harvesting and regeneration resulting in a chess board-like pattern of forest, shrub, and grassland; individual pixels change but the bulk composition remains approximately the same. Fig.B focuses on outlier P2 - the scene where JSS1 indicate only moderate change but JSS indicates large change. The changes in this tile are due to deforestation; bulk class composition changes but not very much but pat-tern changes a lot from forest with shrub patches in 2001 to shrub with forest patches in 2006.