T his paper explores the application of vector agents (V A ), a geom etry-led type of com putational agent, to extract the boundaries and classify real-w orld objects from satellite im ages. T his m ethod has been successfully im plem ented and tested on a m ulti-spectral satellite im age to extract a set of objects w ith real-w orld counterparts. C om parison to the outcom e of an O bject-B ased Im age A nalysis at a single scale of segm entation has also been m ade. The results of the presented approach enables real -w orld phenom ena to be m odelled w ith geographical objects derived intelligently from im ages. T hese objects have geom etric, state and neighbourhood behaviours allow ing them to iterate tow ards realistic output boundaries and robust classification in the sim ulation process.
In the process of determining real world phenomena in remotely sensed images (e.g. an agricultural parcel or a
building), the object boundaries are often interpreted as a set of irregular polygons. These boundaries enclose
homogeneous areas and are collectively able to convert a spatial phenomenon, defined in a continuous environment
(e.g. the remotely sensed image), into a set of objects in a discrete space (e.g. the classified image). However, from a
geographic object-based image analysis (GOBIA) perspective, it may not be easy to determine these regions
precisely. For one, they are extracted without a direct relationship with real-world objects
To tackle these limitations, a vector agent (VA) approach that encapsulates spatial reasoning capabilities (e.g.
use of orientation and size knowledge), will be investigated in order to extract the boundary of real-world
phenomena from a satellite image. Each image object in the model can be interpreted as a level of abstraction of a
real-world object in subsequent iterations of an evolving process from pixel to real-world object. In this regard, this
VA implementation moves toward a classification solution. Hence, the modelled objects are a set of vector agents
that evolve dynamically in a spatio-temporal environment (Figure 1), and once evolved are subject to a single scale.
These VAs provide a dynamic geometry and impose the properties of real-world and simulation environment for the
modelled objects. This is in contrast to previous VA implementations
The initial scenario is formed based on three different classes: agriculture, bare soil and water, found in a subset of an IKONOS image (Figure 2a). First, each desired object is automatically initialized in space. This is done by using the feature space defined by the reflectance information from each of the four spectral bands of the sample image. The feature space exhibits sets of pixels as contiguous clusters with similar spectral reflectance properties. A candidate pixel, which has a minimum spectral distance from the mean of each cluster, is extracted for each class. In addition, a spectral threshold is defined for each class as the maximum Euclidean distance for candidate pixel to belong to the modelled class object. This is equivalent to a minimum Euclidean distance classifier. After initializing, the desired object repeatedly seeks to find such candidate pixels in image space, in an effort to define the full extent of its VA boundary. Also, the other agents are automatically initialized in space, in the same way. Once an agent finds a neighbor pixel that meets the class criteria, this triggers the evolving process of the VA controlled by a set of rules. These rules are defined by structure and geometry primitives to adjust the boundaries of the modelled objects.
An example of the result of this process is shown in Figure 2(b), while Figure 2(c) illustrates the outcome of a single-scale segmentation of the same subset completed with the Trimble eCognition OBIA classification software. The results are promising and indicate that the VA object classification method is capable of evolving towards a modelling of the image in readily classified objects, while traditional OBIA requires the successive and somewhat independents process of segmentation then classification.
This study proposes a new intelligent vector agent approach to extract the boundaries of real-world objects in the image space through spatial reasoning. This approach enables the desired objects to extract their own boundaries based on the existing information for a great variety and complexity of real-world objects. In this study, the different behaviours of desired objects, such as initialising, expanding and shrinking, have been individually investigated. For future development, the full scope of the geometry, state and neighbourhood interactions rules will be explored. Hay, G.J., Castilla, G., Wulder, M.A. and Ruiz, J.R. 2005.An automated object-based approach for the multiscale image segmentation of forest scenes. Int. J. of Applied Earth Observation and Geoinformation, 7, 339–359. Moore, A. 2011. Geographical Vector Agent Based Simulation for Agricultural Land Use Modelling, in Marceau,D.
and Benenson, I. (Eds) Advanced GeoSimulation Models.
Torrens, P., and Benenson, I. 2005. Geographic Automata Systems, International Journal of Geographic
Information Science, vol. 10, no.4, pp.385-412.
Copyright © by the paper's authors. Copying permitted only for private and academic purposes.
In: S. Winter and C. Rizos (Eds.): Research@Locate'14, Canberra, Australia, 07-09 April 2014, published at http://ceur-ws.org R@Locate14 Proceedings 144