The quest for automated land cover change detection using satelite time series data

The paper shows that a Multi-layer Perceptron (MLP) operating on a sliding window of multi-spectral MODIS timeseries data is able to detect simulated change which was introduced by concatenating timeseries of two pixels belonging to two different land cover types. This approach requires training data and is thus supervised. An unsupervised change detection method is proposed based on a non-linear Extended Kalman Filter. The parameters were estimated to represent a triply (amplitude, frequency and phase) modulated cosine function and were separable. Change detection may thus be performed using a classifier operating on these parameters as a function of time.