285
SEGMENTATION OPTIMIZATION FOR AERIAL IMAGES WITH SPATIAL
CONSTRAINTS
Ruedi Boesch, Zuyuan Wang
Swiss Federal Institute for Forest Snow and Landscape Research
8903 Birmensdorf, Switzerland - boesch@wsl.ch, zuyuan.wang@wsl.ch
KEY WORDS: Image Segmentation, Optimization Methods, Parameter Estimation
ABSTRACT:
Unsupervised segmentation methods are important to extract boundary features from large forest vegetation databases. Finding
optimized segmentation algorithms for images with natural vegetation is crucial because of the computational load and the required
reproducibility of results. In this paper, we present an approach how to automatically select optimized parameter values for JSEG
segmentation. The parameter evaluation is based on a spatial comparison between segmented regions and manually acquired ground
truth. City block distance will be used as error metric to define discrepancies between available ground truth and segmentation.
Varying the parameter range of values systematically allows to compute corresponding error areas. The smallest error area represents
the optimized parameter value.Dependent on the lightness distribution of the selected images and the chosen color quantization, the
spatial comparison with the ground truth is limited to local optimization.
1. INTRODUCTION
High resolution images acquired by aircraft or satellites are an
important data source for extracting landscape boundaries and
other vegetation structures. Large scale landscape and forest
inventories depend on reproducible methods for the delineation
of distinct vegetation types. Unsupervised segmentation
methods represent suitable approaches to extract the required
delineation features. Compared to natural images in illustrations
and other man-made contexts, remote sensing data for
landscape inventories contains often vegetation textures with
weak edge structures and faint background.
Swiss National Forest Inventory (NFI) has to record state and
changes of the Swiss forest and makes use of a large image
database with 7000 aerial images, taken at a scale of 1:30000
with a resolution 0.5m for interpretation and area-based analysis.
The image acquisition took place over the last 5 years, therefore
varying illumination and inconsistent color distributions for
vegetation areas are inherent.
2. SEGMENTATION METHODS
Image segmentation techniques can be classified according to
several schemes (Gauch & Hsia, 1992), (Cheng et al., 2001).
They distinguish between statistical methods, edge detection
techniques, split and merge algorithm, methods based on
reflectance models and human color perception. The large
number of segmentation publications over the last decades is
impressive, but selection of the best method is still a difficult
task (Espindola et al., 2006),(Kato et al., 2001). When a suitable
method has been evaluated, the segmentation has to be
parametrized carefully. The ill-defined nature of the
segmentation problem itself does not ease the parameter
selection task. Parameter values of segmentation methods
consist not only of discrete box-filter dimensions, but also of
scaling ranges, e.g. for region growing and color quantization.
Therefore evaluation of the optimal segmentation with the
corresponding optimal parameter set is far from being a solved
problem (Unnikrishnan et al., 2007),(Ndjiki-Nya et al., 2006).
To reduce the evaluation complexity and to achieve robuster
comparability this paper focus on parameter optimization.
2.1. JSEG
Compared to man-made structures, colors for natural landscape
types like forest or agriculture fields contain typically increased
luminance and hue variance, mainly caused by fractional
fluctuation of surface structures and temporal effects. Due to
soft or missing edge structures, local mean values for hue and
luminance are fairly similar, therefore segmentation algorithms
reveal increased over- and under-segmentation effects
(Monteiro & Campilho, 2006). This strongly influences the
segmentation results and emphasizes the importance of a
systematic and quantitative parameter evaluation.
The unsupervised segmentation algorithm JSEG (Deng &
Manjunath, 2001),(Wang et al., 2004) has been selected mainly
due to its robust performance for different vegetation images. In
this paper we focus on the parameter optimization for JSEG to
achieve an automatic delineation of homogenous vegetation
areas (Wang & Boesch, 2007).
Systematic parameter evaluation is often limited by time and
operational constraints and specially for applications with a
broad usage expert values are therefore to be handled with
caution. Often expert values are just derived to deliver
working results. Empirical values often hamper the tedious
search for better settings. But for complex implementations like
JSEG, it is also obvious, that no simplified underlying
mathematical model like a convolution can be assumed. The
peer group filtering of the color quantization, iterative region
growing and seed determination of JSEG cannot be reduced to a
functional description, therefore according to neural network
concepts (Tyukin et al., 2007), optimization of parameter
selection remains an approximation problem with infinite
randomization of parameter values. The published parameters
for JSEG lead to robust results for image collections with
homogeneous background (e.g. Berkeley Segmentation
Database (Martin et al., 2004)), but forest inventory images are
often prone to over-segmentation.