Knowledge-based segmentation of Landsat images

Automatic Spectral Rule-Based Preliminary Mapping of Calibrated Landsat TM and ETM+ Images

“…Blonda well-known problem in the RS literature [1]- [8]. In [1], kernel image information layers are defined as those reliably extracted from RS imagery by means of: 1) domain-specific (e.g., spectral, geometric, textural, semantic, and contextual) prior knowledge and 2) unsupervised (e.g., automatic and data-driven) image processing techniques. This definition implies that kernel image information layers employ no inductive learning-byexample mechanism, i.e., they require no target class sample.…”

“…5) In the proposed implementation, kernel spectral categories are: a) conceived for Landsat-5 TM and 7 ETM+ imagery calibrated into planetary reflectance (albedo) and atsatellite temperature [10]; b) scalable to other existing medium-to high-resolution spaceborne optical sensors sensitive to MS and PAN portions of the electromagnetic spectrum, such as the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) and Systeme Pour l'Observation de la Terre (SPOT-4 and SPOT-5) (refer to Table I, where some acronyms adopted in the rest of this paper are introduced, and see Fig. 1); c) capable of replacing and/or extending the set of kernel categories implemented in the alternative rule-based approaches, such as [1] (namely, water, roads, vegetation, nonvegetation, and ambiguous); d) mutually exclusive and totally exhaustive [11]; e) consistent (in terms of one-to-one or many-to-one relationships) with the set of land cover classes adopted by levels I and II of the U.S. Geological Survey (USGS) land cover classification scheme (refer to Fig. 2, where the land cover hierarchy adopted in this paper is depicted).…”

“…In the Marr sense, a primal sketch of an input image "explicitly reveals information about geometrical distribution and organization of color (intensity) changes" [12, p. 37]. 1 The proposed kernel spectral category detection system can be considered of current interest to a large segment of RS readership owing to its potential application domain, which is threefold. 1) In recent years, the launch of (very) high resolution (VHR) spaceborne MS scanners has meant that, for many applications, purely supervised analysis is no longer feasible.…”

“…To overcome the limitations of purely supervised classification of an RS imagery, two-stage hybrid learning systems are recommended [2], [9, p. 185], [14]. In a hybrid data learning framework, a twostage cascade of the proposed spectral knowledge-based first stage with stratified image enhancement, clustering, 1 Marr, who stood against the incorporation of syntactic and semantic knowledge about the structure of a scene into the early stages of visual processing [13], proposes a famous representational framework for deriving shape information from images. It consists of a hierarchy of a primal sketch (or preliminary map), a two-and-a-half-dimensional (2.5-D) sketch, and a three-dimensional (3-D) model representation.…”

Interpreting Line Drawing Images: A Knowledge Level Perspective

“…Ton et al [9] explicitly based their Landsat image segmentation method on a form of classification known as systematic refinement, though a recent review [10] of aerial image interpretation systems suggests that heuristic classification [3] is more commonly adopted, albeit implicitly, in that domain. Configuration models are intended to describe design tasks in which the goal is to assemble some artifact given a catalogue of available components and a design brief in the form of a set of requirements on and constraints between components.…”