canopy cover or leaf area is an important biophysical param-
eter related to ecological processes such as habitat selection,
evapotranspiration, and carbon cycling. The rise of widely
available very high-resolution imagery has led to a novel ap-
proach to characterize canopy cover and leaf area based on
spatial information. This approach relies upon a relationship
between the pattern of canopy tree crowns/gaps with canopy
cover/leaf area (Wulder
et al
., 1998; Bruniquel
et al
., 1998;
Song and Woodcock, 2003; Song 2007). Several studies have
investigated the link between image spatial information and
canopy cover or leaf area in boreal and temperate regions (e.g.,
Wulder
et al
. 1998; Bruniquel
et al
., 1998; Moskal and Frank-
lin, 2004; Song and Dickinson, 2008; Gray and Song, 2012).
While image texture has been used to classify mangrove spe-
cies (Wang
et al
., 2004), to the best of our knowledge, it has
not been used to determine canopy structure for mangroves.
Field Measurements
The
LAI
measurements were collected near the town of Puerta
Villamil on Isabela Island in the Galapagos Archipelago, Ecua-
dor within mangrove forests consisting of three mangrove species
common in this region:
Rhizophora mangle
(red),
Avicennia ger-
minans
(black),
and Laguncularia racemosa
(white), and as well
as several associate species
Conocarpus erectus
(button or button-
wood mangrove) and
Hippomane mancinella
(manzanillo)
.
Field data were collected during the summer of 2009 at
46 plots. A subset of 14 plots containing only white man-
groves (
Laguncularia racemosa
) were analyzed to control for
the effects of the spectral contribution of other species. While
random sampling is ideal, samples were collected along a
series of transects at regular intervals (e.g., established trails)
with additional targeted sampling to ensure the full-range
of
LAI
was sampled. This sampling protocol was used due to
the logistical constraints of dense vegetation and the require-
ment of non-destructive sampling within the Galapagos
National Park. The sampling/plot locations were selected
a
priori
to image acquisition. Observed
LAI
values ranged from
1.8 to 5.2. Digital hemispherical photographs (
DHP
) were
taken at the center of the 10-meter diameter field plots. A 5
megapixel, Nikon Coolpix 5000 camera and a FC-E8 fisheye
lens with equidistant projection were used to take the
DHP
s.
Additionally, canopy height, substrate conditions, mangrove,
and associated species were recorded at nine points at each
plot. Plot centers were recorded using a Trimble GeoXT
GPS
unit and differentially corrected to a 95 percent horizontal
positional accuracy of less than 1.5 meters. The
DHP
was used
to calculate clumping adjusting
LAI
(referred to as true
LAI
or
tLAI
), using Can-EYE software. Previous studies indicate this
software-hardware configuration provides accurate estimates
of
LAI
in tropical forests (Kraus
et al
., 2009).
QuickBird Image Texture
A QuickBird image was acquired for the study area on 27 Au-
gust 2008. The imagery was geometrically corrected using the
Orthorectification algorithm in
ENVI
which incorporates the
rational polynomial coefficients from the satellite and ground
control points. The root mean square error was found to be
less than 1.5 m using 16 independent
GCPs
. Image texture was
calculated using Grey Level Co-occurrence Matrices (
GLCM
),
a widely used method of quantifying image texture (Haralick,
1973). Seven
GLCM
-based image texture metrics were calculat-
ed using the QuickBird 0.5 m panchromatic imagery with PCI
Geomatica v. 10.3. All
GLCM
texture statistics were calculated
using 16 grey levels and a window size of 11 × 11 pixels (~6.6
m). Texture measures were computed in four directions (i.e.,
0°, 45°, 90°, and 135°) and non-zero values (i.e., valid cal-
culations) were averaged to obtain omni-directional values.
If zero-values were calculated in all four directions, then
that data point was removed from further analysis. Four lag
distances (i.e.,1, 3, 5, and 7 pixels) were tested, producing 28
possible metrics of
GLCM
texture.
Statistical Analysis
The statistical analysis was conducted using
MATLAB
, al-
though the analysis could have been done in other statisti-
cal software packages. Non-parametric Spearman’s ranked
correlations (r
s
) were used to identify significant relationships
without the stringent assumptions of data normality, linear
relationships, or outliers required of a Pearson’s correlation
or Ordinary Least-Squares Regression. Twenty-eight repeated
tests were made comparing image texture to ground-based
LAI
measurements, leading to the multiple comparison problem.
To account for this problem, the Bonferroni Correction and
FDR
were also computed and compared to the p-value from
the Spearman’s correlation to illustrate the potential effect of
the multiple comparison problem when examining multiple
remote sensing products in order to test for a significant rela-
tionship with an observed phenomenon on the ground.
Results and Discussion
Table 1 lists the results of the correlation analysis comparing
image texture and leaf area index. Using the unadjusted p-
value, three models are significant (p <0.05). Simple interpre-
tation on the basis of unadjusted p-values would identify the
Correlation - Lag 7 model as highly significant (p <0.005), and
the best model overall. Using the methods to address the mul-
tiple comparison problem, the results suggest a different inter-
pretation. Using the Bonferroni Correction, none of the tests
meet the adjusted
α
-value of 0.0018 (i.e.,
α
-value = p-value / #
tests = 0.05 / 28; Table 1). Similarly, the q-value computed us-
ing
FDR
also shows that none of the tests meet the threshold,
as the best model has a p-adjusted value of just 0.089. Thus,
while the unadjusted p-value indicates significant results,
when accounting for the multiple comparison problem, none
of the results are found to be significant
.
However, an alternative interpretation is that all three
models with an unadjusted p-value <0.05 had the same lag
distance. In other words, all of the significant models were de-
tecting a pattern at a similar scale. While one or more of these
models may be a false positive, these results indicate that the
lag distance of seven pixels may be a meaningful result, even
if the specific
GLCM
metric may be uncertain. Even though a
strong predictive model cannot be created, using an approach
similar to Gelman
et al
. (2012) of carefully interpreting the
significance in the context of the multiple comparison prob-
lem, the results still indicate the scale at which image texture
is related to
LAI
in this environment and provide the potential
scale for future studies to examine.
Second Case Study: Mapping Trends in Inter-annual NDVI
Amplitude the Sahel and Soudan Regions, Africa, 1982 to 2005
Context
The Sahel and Soudan regions of Africa form a latitudinal cli-
matic and ecological transition from the Sahara Desert in the
north to the humid tropical region of Central Africa. The Sa-
hel and Soudan receive between 250 mm to 500 mm and 500
mm to 1,100 mm of mean annual precipitation, respectively
(FAO, 1998). Since the 1960s, the Sahel region has experi-
enced a substantial downturn in seasonal precipitation that
has resulted in drought and famine in the late 1960s, early
1970s, and again in the early-mid 1980s (Hulme
et al.
, 2001;
Nicholson, 2001; Nicholson, 2005). Recent research based on
satellite remote sensing vegetation monitoring has revealed
that seasonally integrated
NDVI
, a proxy of net primary pro-
duction (NPP), have been increasing in the Sahel and Soudan
since 1982 (Tucker
et al.
, 1991; Tucker and Nicholson, 1999;
Eklundh and Olsson, 2003; Olsson
et al
., 2005; Anyamba
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
December 2015
923
