Abstract
Spatial modelling methods usually utilise pixels and image objects as the fundamental processing unit to address real-world objects (geo-objects) in image space. To do this, both pixel-based and object-based approaches typically employ a linear two-staged workflow of segmentation and classification. Pixel-based methods often segment a classified image to address geo-objects in image space. In contrast, object-based approaches classify a segmented image to determine geo-objects. These methods lack the ability to simultaneously integrate the geometry and theme of geo-objects in image space.
This thesis explores Vector Agents (VA) as an automated and intelligent processing unit to directly address real-world objects in the image space. A VA, is an object that can represent (non)dynamic and (ir)regular vector boundaries (Moore, 2011; Hammam et al., 2007). This aim is achieved by modelling geometry, state, and temporal changes of geo-objects in spatial space.
To reach this aim, we first defined and formulated the main components of the VA, including geometry, state and neighbourhood, and their respective rules in accordance with the properties of raster datasets (e.g. satellite images), as a representation of a geographical space (the Earth). The geometry of the VA was formulated according to a directional planar graph that includes a set of spatial reasoning relationships and geometric operators, in order to implement a set of dynamic geometric behaviours, such as growing, joining or splitting. Transition rules were defined by using a classifier (e.g. Support Vector Machines (SVMs)), a set of image analysis operators (e.g. edge detection, median filter), and the characteristics of the objects in real world. VAs used the transition rules in order to find and update their states in image space. The proximity between VAs was explicitly formulated according to the minimum distance between VAs in image space. These components were then used to model the main elements of our software agent (e.g. geo-objects), namely sensors, effectors, states, rules and strategies. These elements allow a VA to perceive its environment, change its geometry and interact with other VAs to evolve inconsistency together with their thematic meaning. It also enables VAs to adjust their thematic meaning based on changes in their own attributes and those of their neighbours.
We then tested this concept by using the VA to extract geo-objects from different types of raster datasets (e.g. multispectral and hyperspectral images). The results of the VA model confirmed that: (a) The VA is flexible enough to integrate thematic and geometric components of geo-objects in order to extract them directly from image space, and (b) The VA has sufficient capability to be applied in different areas of image analysis. We discuss the limitations of this work and present the possible solutions in the last chapter.