Abstract
We present a fast and statistically principled approach to land cover change detection. A reference statistical distribution is fitted to prior data based on off-line analysis, and an adaptive metric based on the exponentially weighted moving average (EWMA) of normal scores derived from p-values are tracked for new or streaming data, leading to alarms for large or sustained change. Methods which can track the origin of the change are also discussed. The approach is illustrated with a geographic application which involves monitoring remotely sensed data to detect changes in the normalized difference vegetation index (NDVI) in near real-time. We use Wal-Mart store openings as a nontraditional way to monitor and validate known cases of NDVI change. The proposed approach performs well on this validation dataset.