A MULTIRESOLUTION REMOTELY SENSED IMAGE SEGMENTATION METHOD
COMBINING RAINFALLING WATERSHED ALGORITHM AND FAST REGION
MERGING
Min Wang 3
d Key Laboratory of Virtual Geographic Environment (Nanjing Normal University), Ministry of Education, Nanjing,
Jiangsu, 210046, China - sysj0918@126.com
KEY WORDS: Multiresolution, Image, Segmentation, Method, Algorithms
ABSTRACT:
Nowadays object oriented image analysis becomes a hot issue in the field of image processing and interpretation because of its more
robust noise removing ability, more abundant image features and expertise knowledge involved in analysis. The first and most
important step of object oriented image analysis is image segmentation, which segments an image into many visual homogenous
parcels. Based on these parcels, which are ‘objects’ not ‘pixels’, more features can be involved which facilitates the succeeding
image interpretation. In this work, a multi-resolution image segmentation method combining spectral and shape features is designed
and implemented with reference to the basic ideas of eCognition, a famous object oriented image analyzing software package. The
algorithm includes the following steps. 1) The initial segmentation parcels, so called the ‘sub feature units’ are obtained with
rainfalling watershed algorithm for its fast speed and pretty good initial segmentation effects. 2) A fast region merging technique is
designed to merge these sub feature units in a hierarchy way. A scale parameter is used to control the merging process, which stops a
merge when the minimal parcel merging cost exceeds its power. A multi-resolution segmentation can be implemented with different
scale parameters, for smaller scales means less cost while merging which create smaller parcels, and vice versa. Several experiments
on high spatial resolution remotely sensed imagery are carried out to validate our method.
1. INTRODUCTION
Nowadays object oriented image analysis becomes a hot issue
in the field of image process and interpretation. The basic idea
of this kind of method is to segment an image into parcels,
extract features from the parcels, and then complete the whole
image interpretation with classifying the features. The main
advantage of object oriented image analysis lies in that it deals
with parcels, which are ‘objects’, not pixels, which causes more
abundant features and spatial knowledge involved in analysis.
Besides, with more robust pepper noise removing ability, it also
brings more comprehensible interpretation results (Aplin et al.,
1999). eCognition (Definiens, 2007) is a world famous object
oriented image analysis software, in which the multiresolution
image segmentation method (Baatz et al., 2007) is a key and
patented technology, whose technological details hasn’t been
opened to the public yet. In order to implement our object-
oriented image analysis software package for information
extraction from high spatial resolution remotely sensed imagery,
we design and implement a multiresolution image segmentation
method combining spectral and shape features, with reference
to the basic ideas of eCognition. Our method is validated with
several successful experiments on high spatial resolution
remotely sensed imagery. 2
2. METHOD PRINCIPLE AND STEPS
When grouping pixels into very small sub feature units at the
beginning stage of our algorithm, it’s of little use to import
shape feature. In our method, an initial segmentation is firstly
carried out only with spectral features to obtain the sub feature
units. Shape can then be introduced into the algorithm to
control the further merging of these feature units with suitable
size. We use rainfalling watershed algorithm to create these sub
feature units for its fair segmentation precision and very fast
algorithm speed, which is important for processing remotely
sensed imagery commonly with large data volumes. But mainly
due to image noise, most watershed algorithms including
rainfalling watershed have a serious over-segmentation
shortcoming. Sometimes it causes that there exist a large
number of very small parcels scattered in the output
segmentation. A pre or post image processing should be carried
out to remove this adverse influence for further analysis. In our
work, we take the latter one, which deals with these very small
units in a unified region merging way.
2.1 SUB FEATURE UNIT EXTRACTION
Watershed algorithm is a pretty good image segmentation
method based on image grey values. A classical implementation
of watershed is based on immersion simulation [Vincent et al.,
1991]. Watershed segmentation can also be implemented in a so
called rainfalling manner. Its principle is to find a steepest
routine of every pixel on the simulated image topographic
surface, and a watershed base is defined as the pixel set whose
downriver routine ends at a same altitude local minimum. The
algorithm includes two main steps: 1) flooding stage: flood the
image with some altitude threshold to create partial ‘billabongs’
to reduce the high frequency signal parts caused by noises so to
suppress the over-segmentation of common watershed
algorithms; 2) rainfalling stage: in order to classify a pixel
which hasn’t fallen into certain billabong, a rolling down route
of a raindrop on that pixel is simulated, and all the pixels under
this route will be grouped into one class (belong to a same
watershed). After all these pixels are labelled, the segmentation
will be terminated. A critical issue of rainfalling watershed
segmentation implementation lies in correctly dealing with the
local levels embedded in the slopes [Stoev, 2000].