The global probability layer contains a 0-100% cropland probability for each pixel which was derived from a classification tree process using a bagging methodology. A sub-pixel training dataset was compiled from a variety of sources and covered approximately 80% of the earth's land surface, excepting Antarctica. Sources of training data included GeoCover Landsat, UN-FAO AfriCover, USDA NASS, US NLCD, Agriculture and Agri-Food Canada, South Africa State of the Environment and CORINE.

25 individual classification tree algorithms were generated using a bagging methodology, each of which included 20% of the total available training or approximately 16% of the earth's land surface. All trees were applied globally, the 25 results were ranked per-pixel and the median result was used as the final cropland probability. On average, the 25 trees explained 47.9% of the root node deviance with just 4 of the 39 available MODIS metrics accounting for 65.8% of the deviance reduction. The two best metrics were the average NDVI within the 12 warmest 16-day composite periods (of the 23 periods that make up a year), and the average red reflectance in the 12 composite periods having the lowest red reflectance.

Last Updated: 3/11/2010