Learned Atoms from Multispectral Satellite Image Time Series

Kurzfassung

With the on going development of the satellite image time series (SITS), information retrieval in earth monitoring and automatic recognition of changes in time series have become important and challenging tasks. The purpose of the present paper is to determine dictionaries of atoms directly learned from the time series images such that a sparse representation of the images is maintained. The filterbanks learned using gradient descent techniques depend on the frequency band at which the satellite’s sensor captures the scene, i.e., for the visible part of the spectra, the learned filters are close to mean filters over small subregions, whilst, for the infrared part of the spectra, the filters are similar to gradient ones. Furthermore, we aim to express the time changes in SITS in a compact form provided by the newly found sparse representations of the images. The experiments are carried on 4 Landsat images at 30 meters spatial resolution, covering a surface of � 59×51 km2 over the city of Bucharest, Romania, and containing information both from the visible and from the infrared domains.