Detecting land cover changes based on their trajectories in the spectral space

A major problem in detecting land cover change using satellite images is to separate the change signals due to actual land cover changes from those due to various types of noise. The decision rules for separating these two types of change signals are referred to as the spectral trajectories of land cover changes. These trajectories, derived from samples of the two types of signals using a decision tree classifier, were applied to three test images. In general, the trajectory method gave accuracies slightly lower than the best accuracies achieved by thresholding difference images or principal component transformed images. While the thresholding techniques rely on human interpreters to derive appropriate threshold values, land cover change trajectories were applied to calibrated images with little human intervention. Therefore, the trajectory approach has the potential for use in operational land cover change detection. Better accuracies using this method could be achieved by refining the spectral trajectories of land cover change processes using improved global data sets generated by new sensor systems like the Moderate Resolution Imaging Spectroradiometer (MODIS)