Comparing and Synthesizing Different Global Agricultural Land Datasets for Crop Allocation Modeling

Cultivated land has been feeding the world for thousands of years. Only in the last few decades, remote sensing is used to assess and monitor the extent and status of cultivated land. One of the greatest challenges when working with existing land cover datasets is the lack of consistent and reliable data on the location and area intensity of cultivation. By most counts land cover datasets identify cultivated areas as those that encompass cropland and highly managed pasture. Extensive pasture and grazing lands are difficult to distinguish from natural grasslands and thus are usually not identified separately. A number of coarse resolution (1km) global land cover datasets exist but the accuracy and extent of the areas classified as cultivated vary widely. These datasets include: IFPRI’s (International Food Policy Research Institute) extent of cultivated area, which was derived from the Global Land Cover Characterization Database (GLCCD) and is based on 1992/93 AVHRR satellite data; GLC2000 which was derived from 2000 SPOT satellite data; Boston University’s Global Land Cover dataset based on 2000 MODIS data. Each of these datasets includes classes related to cultivated areas but each were derived using different criteria, thresholds, etc... and none of them stands out as fully encompassing the areas across the globe that are characterized by cultivation particularly those characterized by a mosaic of cultivation and other natural land covers. It is thus a challenge for individuals and organizations working with these datasets to find a reliable ‘picture’ of cultivation in one dataset. A preliminary analysis of these datasets performed for the Millennium Assessment (MA) found that the MODIS dataset was severely lacking in its representation of cropland and cropland mosaics. It reported a mere 12.6% of global land cover designated as cropland or cropland mosaic compared to 18.3% for GLC2000 and 27.3% from IFPRI. The regions with the most severe deficits were Latin America and Sub-Saharan Africa which is not surprising since these are areas where mosaic classes are most prevalent in the other datasets. The goal of this paper is to derive an integrated cultivated area dataset based by merging the existing crop land surfaces. We first describe the three global cultivated land datasets, and compare their disagreements. Using actual crop census data as a benchmark, we assessed the difference and accuracy of these three cultivated land surfaces. We then proposed a method to exploit the synergy of existing datasets. Finally we apply our method to develop a new synthesized global cultivated land product and a corresponding confidence estimate of the new cultivated land dataset. This new cultivated land has been used in our spatial allocation model (SPAM).