We propose a time series analysis method to extract condensed and relevant information from noisy and redundant Earth observation time series data. Many time series analysis methods attempt to characterise a time series by reconstructing the original time series with a prior specified model. Such methods, like curve fitting, wavelet transform and harmonic (Fourier) analysis, often come with strong model assumptions and arbitrary parameters which must be manually specified beforehand. Furthermore, a model may apply well to some parts of an image, but fail elsewhere.

Our analysis method adopts a new strategy. The different aspects of the characteristic of a remote sensing time series are represented by a set of statistics (which we call coefficients) selected to correspond to the dynamics of natural system. To ensure the coefficients are robust and generic, statistics are calculated independently by applying statistical models with less complexity on shorter segments within the time series. Specific target problems are then addressed by applying machine learning algorithms or more sophisticated statistical methods using some or all of the coefficients as input data.

We applied our methods to a time series of MODIS 250m 16-days composite images for the whole Australia from years 2000 to 2008. Our experimental results (reported elsewhere in these proceedings) demonstrate that the proposed remote sensing time series analysis scheme is both flexible and powerful. It provides a general platform for tackling a wide range of challenging issues in Earth observation such as change detection, land cover classification and dynamic land cover mapping.