A L2-Norm Regularized Pseudo-Code for Change Analysis in Satellite Image Time Series

Kurzfassung

The continuous progress in the acquisition of high-dimensional information (e.g., satellite image time series, or medical screening) im-plies an efficient characterization of changes that occur in a temporal series of data. A pseudo-encoding technique can be designed to represent the changes between two consecutive moments of time, based on the min-imization of a convex error function which has an analytical solution. The domain transformed feature vectors are grouped into clusters using K-Means. The proposed approach results in a better separation between classes and, thus, in an enhanced characterization of temporal changes. The experiments are done on 5 Landsat multispectral images at 30 me-ters spatial resolution, covering an area of approximately 59 X 51 km2 around Bucharest, Romania.