Self-Organizing Map-based Applications in Remote Sensing

Abstract

Remote sensing involves the collection of information about an object from a distance. Often remote-sensing instruments are mounted onboard an airor space-borne platform and typically record electromagnetic energy in specific wavelength intervals, or bands. The electromagnetic energy recorded over a given area contains information about surfaces reflecting or emitting energy. This information can be used for a variety of applications; for example, remote-sensing image analysis can extract thematic information such as land-cover types (Jensen, 2005). Artificial neural network (ANN) techniques have increasingly been employed in the analysis of remotely-sensed images. ANNs can be advantageous in digital image processing in that no assumption is made about the statistical properties of the images, and they are thus widely applicable to a variety of dataset types. In addition, ANNs learn adaptively through examples and have a high tolerance to noisy or incomplete data (Jensen, 2005). ANN model development can proceed via either supervised or unsupervised means, and if adequate training data are available, supervised training may be readily performed. However, obtaining reliable training data in remote-sensing applications is often problematic (Congalton and Green, 1999), as a remote sensor image typically covers a large area, and only a limited number of training locations can be sampled in the field due to cost, time, personnel requirements, and various other logistical constraints, including potential restrictions on access to the study area. Unsupervised image-processing methods— including unsupervised ANNs—can be of significant utility in such circumstances (Filippi et al., 2009). Unsupervised ANNs are used in situations where the correct outputs may not be known, or if it is desired that the network discover or categorize regularities or features in the training data on its own. There is no teacher signal (Hassoun, 1995). The unsupervised Kohonen self-organizing map (SOM) is a two-layer network, with an input fan-out layer, and an output layer (known as the Kohonen or competitive layer), and the method is based upon competitive learning. The Kohonen layer is comprised of a physical net of neurons located at fixed positions (i.e., intersections in a grid of square meshes). Adjacent neurons are assumed to have a Euclidean distance of unity. The input 14