NEURAL CLASSIFICATION OF REMOTELY-SENSED IMAGES FOR REGIONAL KNOWLEDGE: EMBEDDING ANCILLARY INFORMATION

Sensors mounted on satellite platforms are well-known and useful tools to observe the surface of the Earth and its transformations. The recognition of how and to what extent land covers are modified by human activity (in particular, the automatic mapping of built-up areas and their dynamics); the monitoring of natural resources (extension and condition of forests, crop monitoring, etc.); the occurrence of great natural disasters and the assessment of damages: all these are fields in which the usefulness of remotely-sensed information is fully evident. The series of available satellite images, collected at regular time intervals, covers already more than twenty years. They are in numeric format, therefore suitable for immediate computer processing; yet, to analyze them requires some specifically conceived tools in order to derive synthetic results useful for regional knowledge and management. Classification techniques can be considered among the most relevant tools, both to process longitudinal data (time series of images) or purely radiometric data. Many classification techniques are available, based on various statistical models, or different neural architectures. The major doubt is whether the radiometric information alone is sufficient to produce reliable results: in other words, if it guarantees a good recognition of the various aspects of interest. Unfortunately this is not always the case, and it is often necessary to have recourse to some auxiliary (ancillary) data in order to improve the results. If this is the case, which is the best way to merge into a classification procedure the external a priori knowledge, that can be expressed by data and statistics organized in the most different forms? This paper describes a method to embed in the classification procedure itself the relevant information known a priori. Though in a slightly different form, the method is applicable to both statistical and neural classifiers. A land cover recognition empirical case study, that makes use of a Feedforward Artificial Neural Network, is presented as example