Using Land Surface Phenology for Spatio-temporal Mining of Image Time Series: A Manifesto

PROLEGOMENA: We are in an age of intensive earth observation. Critically needed are tools that will enable efficient and accurate characterization of land surface dynamics by analyzing and extracting the spatio-temporal patterns contained in image time series. As domain scientists, rather than computational scientists, we approach the data with some ideas about what constitutes interestingness in our data. One recurrent interesting theme is the distinction between change from a baseline and variation about a baseline. Baseline characterization is a foundational step in scientifically informed data mining. We seek first to incorporate our domain knowledge in the form of expectations of land surface dynamics, so that we can subsequently identify significant deviations from those expectations (de Beurs and Henebry 2004). Indeed, the motivation of extending occasional observation into intensive monitoring is to detect change, quantify disturbance, and enable prediction. The first NASA workshop on earth science data mining identified anomaly detection as a key characteristic of scientific data mining (Behnke et al. 2000).