ABSTRACT:
This paper describes a new methodology to detect and recognize Object on high resolution multi-spectral images, which involves
successively: (1) Resampling the image according to the size of the object to reduce the data involved in the computation greatly. (2)
Geostatistic method and a local indicator of spatial autocorrelation is used to detect and , more importantly, to locate all the local
clusters with high or low reflectance values which are probably the interesting target objects. In this step, by leveraging both spectral
and spatial information, the algorithm requires little or no input from user, and hence can be readily automated. (3) Finally, identify
objects by extracting the spectral and geometry features in small image block of the original images. The approach is implemented
and tested on lm resolution aerial digital images collected in the east sea of China, and the ships sailing or anchoring on the sea
surface are properly extracted.
1. INTRODUCTION
Object detection has always been one of the most popular
researches in remote sensing science. From an application point
of view, the object detection is usually defined as carrying on
localization and recognition process on remote sensing data for
interesting target. Object detection can be divided into two
different stages, namely target searching and localization
(dynamic question) and recognition and confirmation (static
question) . In the static question, targets are located, needing
further recognition and confirmation about attributes of the
targets. In this stage, efficiency is not very important because of
the accuracy request. However, in the dynamic question, target
position is unknown; therefore, searching time is one of the
most important evaluation criterions.
In recent years, high-resolution satellite and aerial imagery has
recently become a new data source for extraction of small-scale
objects such as vehicles, roads, ship and so on...This paper
concern locating and extracting ships from the high resolution
aerial image. At present, most of the ship detection studies are
based on the Synthetic Aperture Radar (SAR) images and only
a small number of the ship extraction researches uses
panchromatic band images of high-resolution satellites.
Therefore, the usual ship extraction algorithms are actually
detection of light-target on dark-background. There are three
notable aspects of limitation. Firstly, band limitation will cause
a waste of spectral information. Secondly, current algorithms
only concern a small block of data, which indeed includes the
target object, but ignore developing method of searching
potential targets on large imagery. Thirdly, a common and also
very important difficulty in remote location and target
recognition is the low efficiency and long calculation time
when massive data is used. Since existing algorithms can hardly
meet the fast increasing demand for real-time information
management, we try to development a novel algorithm which is
a kind of multi-scale strategy of information extraction based
on geostatical and local cluster analysis.
An increase of use of spatial statistics in the analysis of
remotely sensed data has occurred in the last decade. In
particular; geostatistics offers a broad range of techniques that
allow not only the characterization of multivariate spatial
correlation, but also the spatial decomposition or filtering of
signal values [Goovaerts, 2002]. The approach known as
factorial kriging relies on semivariogram to detect multiple
scales of spatial variability, followed by the decomposition of
spectral values into the corresponding spatial components. This
technique was first used in geochemical exploration to
distinguish large isolated values from group wise anomalies that
consisted of two or more neighbouring values just above the
chemical detection limit.
The LISA (local indicator of spatial autocorrelation) statistic
allows the comparison of an observation (i.e., a single pixel or
small group of pixels) with the surrounding ones, followed by a
test procedure to assess whether this difference is significant or
not. This approach has been used recently to detect spatial
outliers in soil samples [McGrath, Zhang, 2003], while the
LISA has been introduced to quantify the degree of spatial
homogeneity in remotely sensed imagery. The novelty of the
proposed approach lies in the geostatistical filtering of the
image regional background prior to testing the significance of
LISA values through randomization, and the development of
two new statistics to combine test results across multiple
spectral bands.
2. SHIP DETECTION BASED ON GEOSTATISTICAL
AND LOCAL CLUSTER ANALYSIS
The method put forward in this paper is an automatic target
detection process, which capitalizes on both spatial and spectral
bands correlation and does not require any a priori information
on the target spectral signature. The technique does not allow
discrimination between types of anomalies.