Automated global land cover classifications using satellite-based data from the Advanced Very High Resolution Radiometer (AVHRR) and multi-directional data from the Polarized and Directionality of Earth Reflectances (POLDER) instrument

Accurate and repeatable global land cover classifications are an integral component for many studies of global change. Data acquired from polar-orbiting satellites provide the synoptic and consistent information needed to perform these classifications. Current global land cover classifications derived from satellite data have achieved accuracies between 70% and 90% for 12 to 17 land cover types. These products, however, have up to date required the intervention of human analysts at some point in the production of the final results.; Parallel to this, much research aimed at understanding the Bidirectional Reflectance Distribution Function (BRDF) of terrestrial surfaces has indicated that the BRDF may contain additional surface information that could be used to improve land cover classifications. The benefits of this added information for classifications has not been fully explored, particularly at the global scale.; The research presented here extends the decision tree classification methods developed with global data from the Advanced Very High Resolution Radiometer (AVHRR) to make them fully automated. The research also explores the potential contribution of global BRDF data from the POLarization and Directionality of Earth Reflectances (POLDER) instrument for global land cover applications. Automated procedures for global land cover classifications are developed using the C5.0 decision tree classifier and evaluated on 8km AVHRR data. The performances of automated pruning procedures with C5.0 are analyzed, as well as any improvements to the classifications when using new machine learning techniques such as boosting and cross-validation filtering of training data. The success of these automated procedures is also evaluated against two current 1km global land cover classifications from AVHRR. Finally, the same procedure is used to examine the benefits of using BRDF data from POLDER in the classification process.; The research culminates in the production of the first fully-automated global land cover products using AVHRR and POLDER data. Results indicate that is feasible to automate the global land cover classification process without compromising the accuracy and/or stability of the products. (Abstract shortened by UMI.)...