Red and infrared space partitioning for detecting land cover change

One of the algorithms included in producing the change detection product for the Moderate Resolution Imaging Spectroradiometer (MODIS) of NASA´s Earth Observing System (EOS) employs directly partitioning the red/infrared spectral space in order to detect human-induced land cover change. Monthly look-up tables for tall woody vegetation, short herbaceous/woody vegetation, and bare ground have been created to detect spectral migration from one cover type to another. Using training data from the 8 km Pathfinder data set and the University of Maryland´s 1 km global classification, a classification tree algorithm was used to partition the monthly look-up tables for all 12 months across four regions of the globe. The globe was split into northern tropical (0-23.5 degrees north), northern extra tropical (>23.5 degrees north) southern tropical (0-23.5 degrees south) and southern extra tropical (>23.5 degrees south) regions. The classification tree was used to find core areas approaching 100% accuracy for each of the 5 classes. Spectral boundaries between classes were revealed in the tree structure by poor classification accuracies, and were labelled as mixed areas. In this manner the problem of overlaying two full classifications will be avoided as mixed pixels will be labelled as such, and only pixels migrating from one core area to another will be labelled as change. Compensation for phenological changes is built into the algorithm as the look-up tables are made for each month in each region. For example, in the northern extra tropical region winter look-up tables have a much smaller percentage of core areas present in the red-infrared space compared to summer look-up tables, due to seasonal brown-off and snow cover. Results from two test areas reveal the utility of this method, particularly in reducing errors of commission