COMPARISON OF STATISTICAL METHODS AND NEURAL NETWORKS IN
A POST-OBJECT CLASSIFICATION FOR FORESTRY REGISTRATION
Bobo Nordahl, Dr.Ing. student
Department of Surveying and Mapping
Norwegian University of Science and Technology, NTNU
N-7034 Trondheim, NORWAY
E-mail: Bobo.Nordahl(aiko.unit.no
Commission VII, Working Group 3
KEY WORDS: Remote sensing, Forestry, SPOT, Neural Networks, Post-Object Classification
ABSTRACT
This paper presents a project where post-object classification has been used for forest stand registration. The
classification method is divided in two steps where the first is pixel-specific and the second is object-based. In the
pixel-specific classification, statistical and neural network classifications have been compared which is a particular focus
of this paper. The stands from the previous forest inventory were used as objects in the second step. Ancillary data such
as slope, aspect and stand with information such as boundaries and classes from the previous inventory, were combined
with the image data. The project method is designed to prolong the forest inventory intervals rather than replace them
in order to reduce the cost of forest management planning. The conclusion for the low productive areas is that they can
not be classified with these methods. The result from the area which is dominated by medium-high site classes, indicated
that homogeneous stands are better classified with a statistical classifier but a neural network classifier could perform
a better result with heterogenous stands. After the object-based reclassification the differences are smaller. The results
do not indicate that neural network classifier would be useful enough to this kind of forest stand classification.
1. INTRODUCTION scale test have been carried out in [Kolstad 1993]. The
main differences from Swedish and Finnish conditions are
The aim of this study is to investigate the possibility of the hilly topography and the smaller average size of the
using remote sensing images in an alternative forestry — stands in Trgndelag.
inventory. The project focuses on the coniferous forests in
a region in the middle of Norway (Trgndelag), with ^ The project consist of three main parts:
Norway spruce (Picea abies) and Scots pine (Pinus
sylvestris). The initiative to start this project came from
the local forestry authorities who wanted to reduce the
high costs of the traditional inventories. It was presumed
that the traditional stand forestry inventory could not be
replaced but the intervals of 10 years could be prolonged
to 20-25 years (the rotation age is 100 to 150 years).
During these intervals, registrations with remote sensing
images to update the previous forest inventory should be
done.
Several other Nordic tests and projects that used remote
sensing for forestry purposes have been studied. A method
for stand delineation has been developed in the program
named SKOGIS [Hagner 1990 and 1991], which could be
used as a method to establish objects. The inductive
approach using cluster analysis to explore the basic
information in the satellite image is outlined in [Strand
1989]. The effect from varying reflection depending on
topographical and background variations is handled in
[Tharaldsen, Angeloff 1992]. A summary of some Nordic
projects and an implementation of SKOGIS for a small-
542
-New inventory classes have to be established,
based on what kind of forestry information at the
stand level that could be separated in the image
and what kind of information that is needed for
forest management planning?
-What kind of ancillary data to be used has to be
selected. Both to establish objects and to get
information about the topography and other
relevant stand information.
-An interpretation/classification method has to be
established. Several methods for improving
classifications have been tested in research
projects during the last few years and some main
areas are discussed in [Richards 1993]. When
working with forest registration the stands should
be a better classification unit than pixels and
[Lócherbach 1992] summarized several
advantages with object-based interpretation
compared to pixel-based interpretation.
The first two parts are only roughly outlined in this paper
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996