Neural networks ensemble for automatic detection of changes from Cosmo-Skymed SAR images

Remote observations in the optical part of the spectrum are generally used to monitor land cover and its changes. However, atmospheric conditions can seriously degrade the performance of optical sensors, which, furthermore, can only operate in daylight. As a consequence, to meet the requirements of promptness, timeliness and reliability, use of synthetic aperture radar (SAR) must be considered. A crucial step forward in Earth observation has been facilitated by the recent (2011) full availability of SARs on the COSMO-SkyMed (CSK) satellite constellation, operated by the Italian Space Agency (ASI). In fact, the four CSK X-band SAR sensors now in orbit are able to provide images not only at 1 m spatial resolution, but also with a very short revisit time, presently as short as 12 hours, irrespective of cloud cover and light conditions. To take advantage of the unique capabilities of the CSK observing system, adequate exploitation of the information contained in the meter-resolution multi-temporal SAR images is necessary. In particular, the large amount of data contained in each image calls for the development of suitable automatic techniques to manage in near-real time the information on land cover changes which are provided by the SAR observations. This paper presents and discusses a novel change detection method, based on the joint use of different neural networks architectures. It is well known that neural networks (NNs) can be very effective in classifying optical and SAR satellite images. Nevertheless, since the relative novelty of the VHR X-band CSK data, their understanding in this case is still under investigation, and only few studies dealing with the land cover characterization and change detection in CSK images have been carried out.