Use of hyperspectral and multi-angle CHRIS Proba image for land cover maps generation

In the last decades, many satellite missions have marked the history of the optical remote sensing, providing a wide choice of products for every kind of application and investigation. The Proba-1 program can not be compared in importance and budget to very important missions like Landsat and SPOT, but the innovations introduced in the payload design have attracted the attention of principal investigators and scientists since the beginning. The new spectral configuration of its high resolution sensor CHRIS can represent an innovation for the wide family of the optical instruments. In fact CHRIS is one of the experiments, as well as NASA Hyperion, concerning the developing of the expensive hyperspectral technology for the satellite environment. Moreover, the innovative solutions for the autonomous orbit maintaining and navigation, permit to the spacecraft difficult manoeuvres and multi-angular acquisitions during the target overpass. The result is an extraordinary and unique database of hyperspectral and multi-directional images acquired several times during the years over predefined test sites. This research work proposes a complete treatment of the CHRIS products since the radiometric and geometric correction till the classification processing. Phases like destriping, atmospheric corrections, spectral adjacency compensation and ortho-rectification have been performed and developed. Several classification exercises have been proposed with the aim of evaluating the impact of the principal key factors of the Proba mission (hyperspectral, multi-angular and multi-temporal) to the final classification accuracy. The directional anisotropy of the refiectance, as well as its temporal dependence, has been well explained and investigated trough the use of spectral models and radiative equations, but studies and applications based on real satellite high resolution and multi-directional data over lands and water bodies are few or still ongoing. Said that, this research work has been addressed to asses if these additional information, which reflect additional costs in terms of satellite technology and image processing are justified with significant improvements for the land cover production, as one of the most diffused application in the research environment. Once again the neural networks have confirmed their effectiveness for the classification of optical images at high and very high resolution. The new spectral, multi-angular and multi-temporal inputs have been well managed and used as additive information for the decision task, without impacting the design of the classification scheme. On the whole the results have been satisfactory in most cases.