Global Elevation Ancillary Data for Land-use Classification Using Granular Neural Networks

The development of digital global databases containing data such as elevation and soil can greatly simplify and aid in the classification of remotely sensed data to create land-use classes. An efficient method that can simultaneously handle diverse input dimensions can be formed by merging fuzzy logic and neural networks. The so-called granular or fuzzy neural networks are able not only to achieve high classification levels, but at the same time produce compressed and transparent neural network skeletons. Compression results in reduced training times, while transparency is an aid for interpreting the structure of the neural network by translating it into meaningful rules and vice versa. The purpose of this paper is to provide some initial guidelines for the construction of granular neural networks in the remote sensing context, while using global elevation ancillary data within the classification process. Introduction The application of neural networks (NNs) in the classification of satellite imagery is one of the benefits in the quest for establishing more efficient methods. There are some clear benefits over statistical classifiers; most of them stem from the fact that NNs make no assumptions about the underlying distribution of the data. The incorporation of ancillary data in the classification becomes therefore straightforward. Ancillary data can be treated as if there were additional spectral bands (i.e., “stacked – vector” approach) without any conceptual incompatibility issues. Overall, the use of ancillary data is sought for two reasons. First, it provides additional information, usually highly uncorrelated to the spectral information at hand, which can be used to either classify more accurately or separate categories that are otherwise difficult to distinguish. Second, the advances in geographical information systems (GIS) have resulted in the development of many databases, some readily available over the Internet, occasionally at no cost. Geospatial technologies are now well advanced and permit the seamless integration of ancillary data held in a GIS database with any raster satellite data. In a way, we can look at ancillary data as a Global Elevation Ancillary Data for Land-use Classification Using Granular Neural Networks Demetris Stathakis and Ioannis Kanellopoulos plug-in to any classification problem. It is information that significantly increases the dimensionality of our input, and it is readily available. Early work on the classification of multi-source remotely sensed and geographic data using NNs is presented in Benediktsson et al. (1990; 1997). Some shortcomings of NNs have been noted in recent years. A number of parameters have to be set by trial and error, most likely by using an educated guess based on experience rather than formal rules. The refinement of those parameters clearly calls for guidance to establish a best practice in the classification problem using NNs (Kanellopoulos et al., 1997). Neural networks can be very accurate, but in general, the more accurate the output, the more complex and difficult it is to explain the plethora of weights and connections that form the NN skeleton. This characteristic of NNs is often referred to in the literature as the “black box” syndrome (Benitez et al., 1997). Clearly, if we are to improve and more importantly trust them, we need to understand the processes taking place inside the NNs. Research to merge fuzzy set theory (Zadeh, 1965) with standard NNs promised a way of creating transparent nets that we are able to interpret. The term Fuzzy Neural Networks (FNN) has been coined a long time ago (Lee et al., 1975). The same basic formation is also referred to as Granular Neural Networks (GNN) (Pedrydz et al., 2001; Bortolan, 1998). GNNs are neural networks in which the processing is done on information granules rather than individual values and fuzzy set theory is deployed to form the granules (Zadeh, 1997). There are several advantages attributed to GNNs. The first comes from the fact that training granules is faster than training original values because the information is reduced (higher level of abstraction). Furthermore, processing is done on a conceptual rather than a numerical level, which opens the way for linguistic processing (Pedrycz et al., 1999) and computing with words (Zadeh, 1999). Finally, information granulation permits building transparent networks. The neural network skeleton is directly converted into a fuzzy rule set (Zadeh, 1997). There are clear guidelines for converting the rule set to neural network skeleton components and vice versa. Nevertheless, many other parameters remain to be set. Some of these parameters are common to standard NNs (e.g., learning rate) whereas others are new (e.g., fuzzy shape function). PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING J a n ua r y 2008 55 Demetris Stathakis is with the Institute for the Protection and the Security of the Citizen/MARS/FOOD, European Commission – DG Joint Research Centre, Via Enrico Fermi1 1, Ispra (VA), Italy ([email protected]). Ioannis Kanellopoulos is with the Institute for Environment and Sustainability, European Commission – DG Joint Research Centre, Via Enrico Fermi1 1, Ispra (VA), Italy. Photogrammetric Engineering & Remote Sensing Vol. 74, No. 1, January 2008, pp. 55–63. 0099-1112/08/7401–0055/$3.00/0 © 2008 American Society for Photogrammetry and Remote Sensing 05-121.qxd 5/12/07 19:18 Page 55