Aviosar 580 Campaign On The Matera Test Site, Neural Network Approach For Optical And SAR Data Classification

Abstract

The objective of the present work has beep tu state the applicabity of a Neural Network approach to the analysis of Multiwavelength Remote Sensed Images and to verify the effectiveness of the neurd tool with respect to a maximum likelihood statistical one. In order to achieve this goal, it has been constructed an integrated datta-set, composed by a single date Thematic Mapper (TM) geocoded image m d a multiband SAR data. Thc microwave data consist of multipolarised C and X bands airborne SAR imagery,acquired on the Southern Italy Matera test site in the framework Of the AVIOSAR 580 itaban campaign in October 1990. In coincidence with the airborne data an extensive ground truth was collected on the site for near 50 homogeneous agricdturd fields. The ground truth has bee digitized and ruperimpoaed to the the remote sensed images. In this manner the points belonging to the fields of known ground truth have been extracted; some of them have been used as training and some as test set. The selected points haye been analyzed usjng both a MaximumLikelihood classification algorithm and a neural network based approach. As a neural architecture a three-layered feedforward Neural Network, trained with the backprogation algorithm, has been used. The Fe6dtS of the comparkon of the two approaches will be shown and discussed both in terms of statistical properties of training data sets used in the learning phwc and in terms of network characteristics.