Feature Extraction from High-Resolution Remotely Sensed Imagery using Evolutionary Computation
Henrique G. Momm, Greg Easso · InTech eBooks · 2011
Feature extraction, in the context of remote sensing, can be defined as image processing techniques to identify and to classify mutual relationships or mutual meaning between image regions (Baatz et al., 2000). The aggregation of image pixels forming image regions and their relationship to other image regions are interpreted and used as cues in the information retrieval process (Quackenbush, 2004). A common approach is to create hierarchical structures of image regions in which fine-scale image regions constitute portions of other coarse-scale image regions (Niemeyer & Canty, 2001). Feature extraction differs from traditional pixel-based remote sensing image classification algorithms in which each, individual pixel (or pixel vector in the case of images with more than one channel) is individually evaluated and assigned to one class (Lillesand & Kiefer, 2000). The difference between low-level information extraction techniques using traditional pixel-based classification methods and high-level information extracted by a human analyst is often referred to as the “semantic gap” (Smeulders et al., 2000). Human analysts use a complex combination of different image cues such as colour (spectral), texture, shape (geometry of image regions), and context (relationship between image regions). However, human analysis of large areas and multiple images is costly and time consuming (Munyati, 2000). As the volume of available remotely sensed imagery increases by many orders of magnitude, one of the challenges faced by many organizations and institutions is converting large quantities of images into actionable information and intelligence. Because human analysis of large areas and sometimes over multiple periods of time is costly and time consuming, scientists have recognized the importance of developing more sophisticated semi-automated or automated feature extraction techniques to improve the information extraction process. The challenge resides in multifaceted problems where the relationship between image’s regions is too complex to be solved by explicit programming (hard computation) and/or these problems require the system to adapt and evolve when image conditions change. This provision is particularly important in remote sensing applications due to changing factors such as variation in sensor spatial and spectral resolutions, change in environmental conditions between images, and specificity of the feature of interest. The use of stochastic algorithms to address these complex feature extraction problems, are now being investigated as a possible alternative; due to their properties of deriving