In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
633 
SEMIAUTOMATIC CLASSIFICATION OF TREE SPECIES BY MEANS OF MULTI 
TEMPORAL AIRBORNE DIGITAL SENSOR DATA ADS40 
L. T. Waser a ’ *, E. Baltsavias b , C. Ginzier 3 , M. Küchler 3 
a WSL, Landscape Inventories, Swiss Federal Research Institute WSL, 8903 Birmensdorf, Switzerland - (waser, 
ginzler, kuechler)@wsl.ch 
b Institute of Geodesy and Photogrammetry, ETH Zurich, 8093 Zurich, Switzerland - manos@geod.baug.ethz.ch 
Commission VII 
KEY WORDS: Forestry, Ecosystem, Classification, Modelling, Aerial, High resolution, Multi-temporal 
ABSTRACT: 
Temporally frequent, cost-efficient and precise forest information requirements for national forest inventories, monitoring or 
protection tasks have grown over time and will continue to do so in the future. New perspectives are given by the airborne digital 
sensor ADS40, which provides entire image strips with high geometric, radiometric and temporal resolution (every three years for 
entire Switzerland). This study presents an approach for semi-automated tree species classification in different types of forests using 
multi-temporal ADS40-SH40 and ADS40-SH52 images from May and July 2007 and August 2008 to support tasks of the Swiss 
National Forest Inventory. 
Based on image segments seven different tree species were classified by combined logistic regression models using spectral 
variables derived from each of the three different ADS40 images. Additional classification was established combining the May and 
July 2007 imagery. Explanatory variables were derived from each image data set using a step-wise variable selection. 
Classifications were five-fold cross-validated for 230 trees that had been visited in field surveys and detected in the ADS40 images. 
The 7 tree species were therefore classified up to four times providing its spectral variability during the vegetation period. The 
overall accuracies vary between 0.67 and 0.8 and Cohen's kappa values between 0.6 and 0.69 whereas the classification based on the 
May 2007 images performed best. Independent from the sensors and acquisition date of the images lowest accuracies were obtained 
for Acer sp.This study reveals the potential and limits of the ADS40 data to classify tree species and underscores the advantage of a 
multi-temporal classification of deciduous tree species with spectral similarities. 
1. INTRODUCTION 
1.1 General Instructions 
New methods for the extraction of forest attributes from 
airborne remote sensing data have grown over time and will 
continue to do so in the future since exact information on forest 
composition is needed for many environmental, monitoring or 
protection tasks. The present study focuses on the classification 
of tree species using multi-temporal ADS40 imagery and was 
carried out in the framework of the Swiss National Forest 
Inventory (NFI) (Brassel and Lischke, 2001; Brändli, 2010). 
Tree species classification is highly correlated to a large 
number of other forestry attributes (e.g. composition, biomass, 
volume, tree damage etc.) and is an essential index in forest 
studies, inventories, management and other forest applications. 
Several studies have integrated multisensoral data to perform 
tree species classification which lead to better accuracies than 
using only a single data input (St-Onge et al., 2004; Hirschmugl 
et al., 2007; Waser et al., 2008b, Waser et al., 2010) or LiDAR 
(Heinzel et al., 2008; Holmgren et al., 2008; Chubey et al., 
2009). A few studies have incorporated multi-temporal data 
(Key et al. 2001) 
According to Guisan et al. (2004) modem regression 
approaches such as generalized linear models (GLMs) have 
proven particularly useful for modelling the spatial distribution 
of plant species and communities. Kiichler et al. (2004) show 
that spatially explicit predictive modelling of vegetation using 
remotely sensed data can be used to construct current 
vegetation cover using information on the relations between 
current vegetation structure and various environmental 
attributes. Thus, logistic regression models seem also promising 
for modelling tree species when analyzing the relationship 
between categorical dependent variables (e.g. tree species) and 
explanatory variables derived from remotely sensed data 
(Waser et al., 2008a and 2008b). 
The objective of this study was to classify semi-automatically 
three deciduous and four coniferous tree species. The 
contribution of three multi-temporal ADS40 images was tested 
and best image combination for tree species classification was 
assessed. Preliminary results are very promising for future 
monitoring, updating and management tasks of a continuous 
Swiss National Forest Inventory (NFI). 
2. MATERIAL 
2.1 Study area 
The study area is characterized by open and closed mixed 
forests and is located in the Swiss central Plateau (approx. 
47°22’ N / 8°28’ E) and has an extent of approx. 7.5 km 1 2 . The 
altitude ranges from 450 m to 850 m a.s.l. The forest area 
covers approx. 5.1 km 2 , and is mostly characterized by mixed 
forest. The dominating deciduous tree species are Fagus 
sylvatica and Fraxinus excelsior and less frequently Acer sp. 
The main coniferous trees are Abies alba, Larix decidua, Picea 
abies and Pinus sylvestris. 
* Corresponding author