Long-term earth observation data, specifically Landsat imagery, offer the possibility of quantifying deforestation rates and the resultant land uses that replace cleared natural forests. Our ability to track the evolution of human economic activity on tropical forest landscapes has improved due to recent advances in 1) data policies making Landsat data freely available, 2) advanced high performance computing and 3) methods for accurate processing and characterization of land cover and land use. The most consistent record of Landsat imagery starts in the early 1980s with the Landsat 4 and 5 Thematic Mapper sensors through the currently operating Landsat 7 Enhanced Thematic Mapper Plus and Landsat 8 Operational Land Imager. The main objective of the proposed activity is to map South American agricultural evolution by commodity during this period of data collection. As cropland, pasture and forestry land uses expand, natural land covers or antecedent agricultural or other land uses, are converted. Accurate measurement of the spatio-temporal trends of these conversions is a critical input to analyses of domestic and international commodity flows and their respective drivers.
Starting in ~1985, we will map natural land cover and human land use at a 30m spatial resolution on an annual basis. Natural land covers will include themes such as humid tropical forest, dry tropical forest and woodland, and grasslands. Wetland status will also be defined as a generic overlay. For human land use, we will focus on commodity croplands including soybean and other crops, pasture for beef production, forestry, and palm oil. Land associated with each commodity will be characterized and a matrix of conversion covering the from-to dynamics created for the 30+ year study period.
Our approach to Landsat data processing is well-established. We exhaustively query and process the entire Landsat archive in order to create a multi-temporal feature space. This feature space is related to a set of training data using decision tree algorithms. Our methods are tuned for large