Using unsupervised neural network approach to improve classification of satellite images

Image classification is an essential process for satellite image processing. It is especially useful for mapping and assessing change in the spatial extent of the different regions over time. Several techniques for processing satellite images allow the use of data provided by the sensors for identifying different land cover classes, such as agriculture, water and urban areas. Among these techniques for extracting knowledge, the authors use neuronal methods. These are applied in various fields ranging from decision support or approximation to the planning, fields of pattern recognition and classification. Consequently, an unsupervised neural networks approach in the satellite imagery field is considered here, which is known as the topological map of Kohonen. The authors apply this method to perform a classification of satellite images. It has a set of tests to allow the determination of appropriate parameters that characterise the Kohonen map. This method was evaluated to obtain optimal classes.