2004
s 296
ha
olor
t/ red
ically
atures
m the
'alick,
ment,
el co-
pixel
image
ration
ection
ed for
atures
konos
filter
set of
omain
annel
using
n the
he
local
pecies
konos
> (rec
xel of
ree IS
ceeds
saved,
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
otherwise coniferous. The Forestime averaging tool then
produces a segmentwise feature proportional to the percentage
of broad-leaved trees as its output.
3.2 Feature selection
To get an idea of best feature set for the parameter estimation
models, the feature data was examined by calculating the
mutual correlation of forest variables and the features, and by
creating estimation models with stepwise linear regression of
SPSS statistical software package. The models obtained with
SPSS contained often five features or more. However, the
marginal utility obtained by adding features above five was
very small with the SPSS regression models, and hardly ever
increased the accuracy of the forest variable estimates obtained
with Forestime. Consequently the feature selection process
concentrated on selecting features best explaining the target
variable variance, and the feature sets used in the estimation
contained six features maximum. In many cases the best
features for separate target variables coincided, resulting in
fairly similar input feature sets (the spectral channels
augmented with the Haralick entropy are a good compromise to
estimate most of the selected variables).
3.3 Segmentation
The basic component of the Ikonos image processing is the
segmentation of the image into homogenous areas (segments),
in the sense of tree species, stem number, tree height and other
relevant forest parameters. The criteria for homogeneity is
based on multispectral radiance or reflectance values and
texture measures. The method allows segments small enough to
include only a single tree.
The multispectral segmentation is based on the method
developed at VTT (Parmes1992), derived from the
segmentation approach presented by Narendra & Goldberg
(1980). The method uses gradient images calculated from the
original image spectral bands as inputs, and links the pixels
belonging to the same image segment with the directed trees
method by Narendra & Goldberg (1980).
3.4 Variable estimation
The segmentation is used in the next step to obtain stadwise
averages from the selected features. The resulting average
images are sampled with uniform random grid to produce a set
of input data vectors. The data set is clustered by K-means
algorithm into a pre-selected number of clusters. Each obtained
cluster is then assigned with the ground reference vector of
target variables as an average of the plot samples belonging to
this cluster. The estimates are produced as weighted sums of the
input sample class probabilities (Häme, et al. 2000, Häme et al.
2001). The estimation results are obtained as raster images, or
in vector data format. The Forestime system principle is
depicted in Figure 4.
337
Average texture Average stem Average species
images number image proportions
mage
Avergae tree
reflectance image
Simple user Calibrated IKONOS
interface image
1 -- - - — " — o— — - - . .=>“Q. U... -—— - —— ) — mm an -— a -
0 = 1
I Feature Tree location I
I manipulation tool '
ub i
| Tree location I
| :
| 1
I 1
1 Segmentation Texture tool pecies !
1 fool proportions 7) I
1 ~—___tool I
pi 1
Segmented Species !
- image proportions I
mage !
1
I
I
I
I
1
I
I
I
i
i
I
|
|
|
1 Segmentwise
I
I
i
I
1
I
I
| Model-based
| estimation Proba model Ground Data
I tool
I
1
I
—»! :
1
r…----
ras2mif ee Resuit Files
conversion
Figure 4 Forestime estimation process
FORESTIME SOFTWARE
The first version of a modular software tool was developed
(Forestime v. 1.0) in the study. The guidelines in the software
design were modularity, development and utilization of in-
house software, and making the method for forest mapping of
large forest areas automatic. The modularity of the system has
several benefits. It allows an easy addition or replacement of
e.g. some feature extraction module with another. The output of
the system might as well be the land cover classification instead
of parameter estimates. The system is not restricted to any
specific imagery, but can fairly easily be configured to use data
from a different type of instrument, or even from several
instruments. The software also includes an API (Application
Program Interface) for integration with other systems like GIS
databases.
The system is a multi-tier application (Figure 5). The Forestime
is running in Windows operating system. The user interface for
the System is provided by a designated GUI or external
software via System’s application program interface. The server
communicates with analysis tools via command line calls. The
Forestime software was realised as object oriented design,
which makes the system easily configurable and expandable. In
software development Rational Rose was used for UML
modeling, and JBuilder for Java implementation. The version
control system was MS Visual SourceSafe. All the analysis
tools were written in C.
