Global land cover classifications at 8 km spatial resolution: The use of training data derived from Landsat imagery in decision tree classifiers

Abstract
This paper reports a study which aims to (i) develop methodologies for global land cover classifications that are objective, reproducible and feasible to implement as new satellite data become available in the future and (ii) provide a global land cover classification product based on the National Aeronautics and Space Administration/National Oceanic and Atmospheric Administration Pathfinder Land (PAL) data that can be used in global change research. The spatial resolution for the land cover classification is 8 km, intermediate between our previously published coarse one degree by one degree spatial resolution and the 1 km global land cover product being developed under the auspices of the International Geosphere Biosphere Program. We first derive a global network of training sites from Landsat imagery, using 156 Landsat scenes mostly from the Multispectral Scanner System, to identify over 9000 pixels in the PAL data where we have high confidence that the labelled cover type occurs. We then use the training data to test a number of metrics that describe the temporal dynamics of vegetation over an annual cycle for potential use as input variables to a global land cover classification. The tested metrics are based on: (i) the ratio between surface temperature and Normalized Difference Vegetation Index (NDVI); (ii) seasonal metrics derived from the NDVI temporal profile, such as length of growing season; (iii) a rule-based approach that determines cover type through a series of hierarchical trees based on surface temperature and NDVI values; and (iv) annual mean, maximum, minimum and amplitude values for all optical and thermal channels in the Advanced Very High Resolution Radiometer (AVHRR) (PAL) data. Highest mean class accuracies from a decision tree classifier were obtained using the annual mean, maximum, minimum, and amplitude values for all AVHRR bands. Finally, we apply these metrics to 1984 PAL data at 8 km resolution to derive a global land cover classification product using a decision tree classifier. The classification has an overall accuracy between 81.4 and 90.3%. The Landsat images used for deriving the training data and the methodology for classification of AVHRR data at 8km resolution can also be applied to 1 km AVHRR data and, in the future, Moderate Resolution Imaging Spectroradiometer (MODIS) data at 250 and 500m resolution. Digital versions of the land cover dataset and detailed documentation can be found on the World Wide Web at http://www.geog.umd.edu/landcover/8km-map.html.