In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
The recent technical progress in the development of digital 
aerial cameras laid the foundation for approaches to distinguish 
tree species out of digital aerial image data. Neubert (2006) and 
Tiede et al. (2009) for ex. describe first approaches for the 
assessment of tree species composition by the use of digital 
aerial photographs. The Bavarian Office for Surveying and 
Geographic Information (Landesamt fur Vermessung und 
Geoinformation (LVG)) conducts flight campaigns since 2008 
on the basis of the latest digital camera systems. Thus, every 
year about one third of Bavaria’s surface is covered by high 
resolution aerial photographs. 
Against this background, the Bavarian Ministry of Food, 
Agriculture and Forestry financed a project of the Bavarian 
Institute of Forestry in cooperation with the enterprise 
GeoCreativ in order to investigate the potential of the actual 
digital aerial images for the semi-automatic assessment of 
Spruce. The project areas Eltmann and Gerolzhofen share more 
or less the same climatic conditions like the regions of western 
Middle Franconia and Lower Franconia that were struck by 
wide-spread losses of Spruce in 2005/2006. Even in this time, 
the observed phenomenon have been supposed to be the 
precursors of a development that would consider vast areas, 
caused by climate change (Ammer et al. 2006). 
1.2 Objectives 
Because of climate change, within short time an overview of 
Spruce stands that are in need of urgent conversion has to be 
achieved. Therefore, reliable, actual information about the 
dispersion of Spruce stands has to be created, using aerial 
photographs furnished by the LVG. In this project, by the help 
of modem software for image analysis and object-based 
classification, vast areas were supposed to be semi- 
automatically analyzed efficiently and in due time. At the end, a 
GIS-layer locating “Spruce” and “other coniferous trees” was 
planned to be available for the project regions. 
2. BASIC INFORMATIONS 
The project regions comprised of two separate areas in the 
„Steigerwald“ (Bavarian forest growth area 5.2 Steigerwald). 
One area contained the communal forest of Eltmann (ca. 2.500 
ha), the second area is located within the „Biirgerwald“ 
Gerolzhofen (ca. 300 ha). Within the project areas grow 
intensively mixed deciduous and coniferous forests on vividly 
undulated terrain of a great variety of expositions. The Bavarian 
Office for Surveying and Geographic Information furnished 4- 
channel aerial photographs (color depth 16 bit, tiff-files; 
original data sets as well as ortho-rectified images). The images 
have been taken by the matrix camera Vexcel Ultracam-X. 
Ground sample distance (GSD) for the research area was 20 cm. 
The spectral bands green, blue, red and near infra-red are 
available in separated form. The flight campaign took place in 
August 31 st , 2008. 3 
3. METHODOLOGY 
3.1 Pre-processing of image data 
The original image tiles show great differences within their 
spectral bands, with the infra-red band possessing the least 
differences between the images. The transfer of classification 
criteria from one original image to another seemed thus to be 
impossible. Furthermore, the individual images partially show a 
distinct color drop especially at forest edges. 
For the preparation of the supplied picture material, it was 
therefore necessary to develop a method for adjusting the four 
spectral channels between the individual images, to extend their 
histograms and to calculate new channels. For this purpose 
ERDAS Imagine 9.3 was used. The different algorithms have 
been combined with the help of the "Spatial Modeler" to an 
overall model. Thus, the enhancement of the individual images 
in each case was feasible in one single step. 
The following processes have been conducted: 
Step 1: Compensation of chromatic heterogeneity between 
images and within individual image tiles. 
The color heterogeneity between the individual images could be 
largely offset by adjusting the individual color channels. 
Exposure differences within the image tiles, particularly at the 
edges of forest areas however, could be corrected only to a 
limited extent. 
Step 2: Enhancement of spectral differences between individual 
trees. 
With the exception of forest edges with strong illumination 
decline, a very significant increase in visual recognition of the 
individual conifer species was achieved on the basis of their 
different colors when displayed on the screen. 
The advantage of this method of image processing is a 
significant time reduction of the effort for the determination of 
training trees. 
3.2 Segmentation / Classification 
To classify coniferous forests covering a great area as training 
samples, Spruce trees and other coniferous trees were used. In 
this operation a probability filter has been applied, that separates 
coniferous and deciduous trees. 
By the use of AOI's (Area of Interest) for Spruce and other 
coniferous tree species, 8 to 14 training trees in the best possible 
allocation within the individual image tile were assigned. With 
increasing heterogeneity of the picture, a larger amount of 
training trees per target class was required. 
The basic settings were chosen in a way that as large as possible 
crown areas were recorded for each segment. The collection of 
several small-crown trees of the same species in a segment was 
deliberately taken into account, in order to improve the 
formation of the segment averages of the spectral bands as far 
as possible. The resulting polygons were attributed by their 
mean values and probabilities with which each segment of a 
species has been classified. Using a probability filter, a pre 
selection of segments for all aspired classes was then conducted. 
3.3 Post-processing and visualization in a GIS 
Fine adjustment, post-processing and visualization of the 
classification results have been accomplished by ArcGIS 9.3 by 
ESRI. For this reason, the generated shapefiles of the segments 
were imported into ArcGIS. A pre-selection was achieved by 
the use of mean values of the reflectance (exclusion of shadows, 
etc.). In a next step, the generated probability limits were 
adjusted to enlarge the target class, while limiting 
misclassifications to a modest range. In this process, the 
necessary border values of the probabilities determined by 
ERDAS Imagine Objective varied considerably. 
For typical training trees, a probability of 60% already proved 
to be sufficient for a reliable classification. For training trees 
with atypical radiation spectrum (e.g. along forest edges, 
damaged trees) a probability threshold of 90% has to be chosen