A Combination of a symbolic and a natural classifier applied to
remote sensing images
Case of study: Penang Island
Lionel Beauge, Alain Ketterlin* Ahamad Tajudin Khader, Ruslan Rainis*, Jersy Korczak*
School of Computer Sciences and *School of Humanities
Universiti Sains Malaysia
11800 USM Pulau Pinang, Malaysia
&
*LSIIT
Department d'Informatique de l'Universite & Louis Pasteur
67084 Strasbourg, France
E-mail : lionel@cs.usm.my
Abstract
In remote sensing, images become more and more complex, stemming from the appearance of higher performance satellites. For years, several kinds of formal models have been applied to analyze them, but complex formalisms lead to high computational complexity e.g. morphology mathematical processings. The ongoing research focuses on a hybrid classifier which can discover automatically structured objects on images, in order to generate thematic maps. In this paper, two unsupervised models are presented. The models are based respectively on conceptual classification (the Cobweb algorithm) and on competitive neural network learning to validate our approach. The case of study is Penang Island which provides several kinds of landscape. The resulting maps are quite similar. Our results are also compared with the thematic map of an expert. Our research tend finally toward an hybrid model which integrates symbolic and neural learning, geometrical aspects and expert knowledge. Our main goal is to produce a map updating system whose maps can be used in the studies on environment monitoring e.g. land and urban utilization, vegetation distribution and its changes, regional development etc.
1. Introduction
Remote sensing images provide spectral values about the region under observation. These spectral data allow to distinguish different spectral forms in images by combining all the bands. But images are mote than a simple collection of spectral values. Pixels which are the basic units of the image are also spatially organized. In fact, spectral and spatial values are two kinds of sources of data which can be analysed in different phases or levels and can be combined to obtain the best classification. The goal of classification is to automatically create thematic maps of the region under observation.
2. Methods
2.1 Segmentation principle
Segmentation consists in finding regions or objects that present a thematic meaning regarding to the expert. It is closed to clustering by the fact that it analyses each pixel-element of the image to form a set of distinct classes. In remote sensing, clustering consists in finding a set of classes in the measurement space é = {C1 ,……., Ck) , where C represents the set of all pixels. Each class Ci i e {I,……k} represents a specific subset of the space of spectral a specific subset of the space of spectral values. Objects in the same classes (" : Oj e Ci) have to be similar and objects from different classes have to be sufficiently different. The problem is to find regularities to the measurements space and to produce homogeneous classes. After segmentation, each pixel is assigned to one class only. The set of classes can be organized as a hierarchy or a single set.
2.2 Unsupervised model
Due to the complexity of images and the missing classified data, most of processings tend to be unsupervised. Knowledge is not enough to give information on each pixel of the image.
In this paper, we present two different kinds of classifier which are both unsupervised. The first one is symbolic and takes spectral and based on neural networks and makes at present segmentation based on spectral values only.
3. A conceptual Classifier
3.1 Principles
Our first clustering method is inspired by research in machine learning. The Cobweb algorithm is a generic clustering algorithm which can handle any kind of vector or composite data [2]. Even though it shares several aspects with "classical" statistical algorithms in the context of pixel-based clustering, it also exhibits several distinctive features which make it especially applicable to satellite remote-sensing images analysis.
The first major advantage of Cobweb is that it produces a hierarchy of classes, sorted form the most general to the most specific. This allows a convenient exploration of the classification results, as will be illustrated in our case study. Eventually, it allows the analyst to generate several maps from the same classification result. The second major advantage of this algorithm is that it is incremental, processing one pixel after the other. It is thus able to process very large images without requiring a lot of memory, since it does not need to keep the whole image in memory.
