BRIDGING REMOTE SENSING ANALYSTS AND DECISION-MAKERS; 
THE SUPPORT OF AGGREGATE-MOSAIC THEORY TO MONITOR TROPICAL 
DEFORESTATION 
M.H. Obbink *', M. Molenaar ^, J.G.P.W. Clevers“, M. Loos“ A. de Gier ? 
"Wageningen University, Centre for Geo-Information, P.O. Box 47. 6700 AA Wageningen, The Netherlands 
— (marion.obbink, jan.clevers, mark.loos)@wur.nl 
? International Institute for Geo-Information Sciences and Earth Observation (ITC), P.O. Box 6, 7500 AA Enschede. The 
Netherlands — (molenaar, degier)(@ite.n! 
KEYWORDS: Theory, Segmentation, Forestry, Land Cover, Mapping, Vegetation heterogeneity 
ABSTRACT: 
The land cover mosaic (LCM) classification concept, which is based on the aggregate-mosaic theory, improves geo-information on 
tropical deforestation for decision-makers. Land cover mosaics are 
spatial units constituting mixtures of different land cover types. 
They are defined by two parameters, that is the mix of different land cover types and the spatial size of these land cover types. In this 
paper, a sensitivity analysis was performed to test the impact of the parameter spatial size on LCM classification results using two 
Landsat TM images of a peatswamp forest in Kalimantan, Indonesia. Five methods were selected to evaluate the LCM classification 
results, the standard remote sensing accuracy method KHAT, and four Landscape Pattern Metrics as applied in landscape ecology. 
Results showed that for spatial sizes up to 150 ha, the total area of forest cover remained constant for both the 1990 image and the 
more fragmented 1996 image. This finding is very useful for assessing the area of forest cover, which often differ widely between 
various sources. From a decision-making point of view it is important that maps are produced with identical parameter settings, 
when comparing temporally different images, because spatial size has effect on the spatial arrangement of forest cover. Combining 
results of KHAT and Landscape Pattern Metrics thematic classes causing significant differences could be indicated. Finally. based 
on variations in LCM classifications, it can be concluded that between 1990 and 1996 forest was depleted not due to logging 
practices, but due to agricultural practices. Such a finding could be very useful to support planning and development strategies, and 
to improve governmental policies to manage tropical rainforests in a sustainable way. 
1. INTRODUCTION 
Global studies of tropical deforestation are based on 
information sources compiled at regional levels. Parts of these 
regional information sources are based on digital maps derived 
from remote sensing data, where forest cover information is 
classified and stored at pixel level. These maps are static and, 
without further interpretation, they do not have much 
explanatory or predictive power about the driving forces and 
mechanisms behind forest cover changes (the actors). Decision - 
makers, however, need such explanatory information on actors 
to decide when and where to take action concerning 
undesirable developments in tropical rainforest areas. 
Consequently, remote sensing image analysts should expand 
their monitoring task of land cover at pixel level towards 
identification of actors, design models on actors behavior, and 
prepare scenario's on the impact of the actors. This calls for an 
underlying remote sensing theory that describes how to retrieve 
such explanatory information form remote sensing data. An 
important issue of such a theory is handling the problem that 
vegetation in tropical rainforest areas is extremely 
heterogeneous (Whitmore, 1998). It often consists of different 
mixtures of trees, shrubs and grasses at different spatial 
aggregation levels resulting in different vegetation structures. 
Consequently, vegetation heterogeneity should be expressed as 
variations in vegetation composition and vegetation structure at 
different spatial aggregation levels. As such, variations in land 
cover mosaics can be related to different actors. 
A range of innovative processing techniques have been 
developed to improve per-pixel based classifications. Howe ver, 
the majority of the techniques either address vegetation mixture 
  
(to reduce autocorrelation or spectral overlap) or vegetation 
structure (to handle image texture). Examples of reducing 
autocorrelation are segmentation algorithms (Hill, 1999), 
filterin g (Palubinskas et al., 1995), and Markov random fields 
(Cortijo & Perez de la Blanca, 1998). Examples of reducing 
spectral overlap are co-occurrencies (Kushwada et al.. 1994), 
semi-variograms (Woodcock & Strahler, 1987), and cover- 
frequencies (Bandibas et al. 1995). Examples of handling 
image texture are wavelet transforms (Carvalho et al., 2001: 
multiresolution segmentation (Burnett & Blaschke, 2003), and 
post-classification aggregation (Beurden van & Douven, 1999). 
Aggregate-mosaic theory describes the concept of land cover 
mosaic (LCM) classification to address both vegetation mixture 
and vegetation structure at higher aggregation levels than the 
pixel when classifying remote sensing data (Obbink, et al., 
2002). Land cover mosaics are spatial units constituting 
mixtures of different land cover types. They are defined by two 
parameters, that is the mix of different land cover types and the 
spatial size (expressed as minimum -area) of these land cover 
types. Based on expert knowledge, variations in land cover 
mosaics can be related to different actors. The objective of this 
paper is to demonstrate the LCM classification concept for two 
temporal images of a peatswamp forest in Kalimantan, 
Indonesia. Specifically, a sensitivity analysis was performed to 
test the impact of the parameter spatial size on LCM 
classification results, and on variability and arrangement of 
forest cover and forest cover pattern using KHAT and four 
Landscape Pattern Indices. 
* Corresponding author. Tel: +31-317 474 721; Fax: *31-317 419 000 
790 
Interne 
2; 
2.1 
Th 
SW 
Ba 
  
30/ 
are 
bot 
gec 
to : 
aca 
clas 
stat 
Lai 
  
Fig 
199