In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
409
Gray level co-occurrence matrix (GLCM) based texture
LH™CM£)= y iP u
_
F
2. HaiKaQm*M = / i ■■■ '
¿U l+O-/) 1
I
3, « GMtnut CC©3 = y i P 4i i;i-i> z
4» Sl^Tsdsfiil intfiwcCSnfl = vT"*
when VA == y 4 (i - M&3 5
*■*
LifemmilsfifjCSl) = y i ji—j|
ar-ji
& Emlrmpy&N} = y 4 (-1*5 Fj)
¿4*
L4SM3 = y i
Ltevn« Difference C/D) =
.v-i
9. GLDT J^iacr 5#iKmd Mswien* IS ASM) — X ly
lb*6
JKF-S
10, t*aropy i££N} = y V*. <-&* l\)
lb«P
P (i, j) is the normalized co-occurrence matrix such that
SUM (i, j-0, N-l) (P (i,j)) = l.
V(k) is the normalized grey level difference vectorV(k)
: SUM (i, j=0, N-l and [i-j ] = k) P (i, j)
Sum & difference histogram (SADH) based texture
parameter
r jr
1, Mean <j4 = —=—-
I
|2 -;! 5
, , £* K ~*M
2»Mean dtvuitiim (Jit?) =
n
¡E~0^ -jy 3
3kM#sf5 Euclidean dist&n-cm CM ED} = j
X (*a -p) 3
*y
A.Variance (ir) =
n, -1
y
i.sk**-nssHSk)=
X t* w - j*) 4
IKurtosis (Eu) = —-————
(« - l>r*
Y, .'V*.
8. fn#r#y<r) = *"*** «
*
Z x
X*.* i / m
4* = pixel value of pixel (*■*.§) in kernel, ' * = the number
of pixels that is summed, X c - the kernel’s center pixel
value, P* = the normalized pixel value.
Table 1. Formulae of texture measurements used in this study
9. Entropy Qi) = — y p u i»£p JL wrer* ja 4
4. RESULTS AND ANALYSIS
The field biomass data from the 50 field plots ranged from
52t/ha to 530t/ha. In all modeling processes, the 50 field plots
were used as the dependent variable and parameters (AVNIR-2
and/or SPOT-5) derived from different processing steps were
used as independent variables.
The best estimates of biomass using simple spectral bands of
AVNIR-2 and SPOT-5 as well as different combinations of
band ratios and PCA produced only ca. 60% useable accuracy
due to (i) the complexity of forest structure and terrain in the
study areas, (ii) The very high field biomass in this study area
(52t/ha to 530t/ha), and (iii) strong multicollinearity effects
among the 8 bands and band ratios from the two sensors used.
A notable improvement was observed for both sensors using
texture parameters (Table 2). For single band texture, the
highest (ANVIR F= 0.742 and SPOT-5 r 2 = 0.769) and lowest
(ANVIR r 2 =0.309 and SPOT-5 1^=0.326) accuracies were
obtained from the texture parameters of NIR and Red bands
respectively. The pattern of accuracy was similar to that
obtained using raw spectral bands although the performance
was much higher for texture measurement. Moreover, as with
raw data, the second highest accuracies (ANVIR r 2 =0.547 and
SPOT-5 1^=0.615) were also obtained from green and SWIR
bands using AVNIR-2 and SPOT-5 data respectively. These
patterns of improvement were consistent for both sensors and
very much in agreement with the general behavior of interaction
between different wavelengths and vegetation. Thus we found
that texture measurement enhanced biomass estimation across
all bands but greater improvement was observed from the bands
where reflectance from vegetation is higher.
However, unlike raw spectral bands and simple ratios of raw
spectral bands, texture parameters from all bands together
(either all bands of AVNIR-2 or SPOT-5) were found to be very
useful, and obtained accuracies of 0.786 (r 2 for AVNIR-2)
(model 1 in Table 2) and 0.854 (r 2 for SPOT-5) (model 2 in
Table 2) Apart from the improved accuracies the developed
models (using all texture parameters of an individual sensor
together) were significant and no multicollinearity effects were
evident.
When texture parameters from both sensors were combined
together in the model (model 3 in Table 2), as well as all texture
parameters of PCA of both sensors together (model 4 in Table
2), and all texture parameters from averaging of both sensors
together (model 5 in Table 2), very significant improvements
were obtained although PCA was not found to be very effective.
The highest (r 2 =0.91) and the second highest (r^O.90)
accuracies were obtained from the texture parameters from the
averaging of both sensors, and texture parameters of both
sensors in the model respectively. These differences were
attributed to the fact that averaging is a type of data fusion, and