Bayesian Updating of Land Cover Classification in th Brazilian Provinces of Amazonas and Mato Grosso

English Abstract:

Land cover classifications are extremely useful tools that can be applied to help solve issues in policy, social sciences, conservation and more. However, they only capture information on a region for a snapshot in time. Given the dynamic state of the Earth and the impact of anthropogenic changes to land use and land cover on various timescales, there is need for continuous tracking of landscape change. This is of particular relevance for areas such as the Brazilian Amazon, where the rate of landscape change, especially from deforestation, has fluctuated over the past few decades. We have developed an algorithm in R (Bayesian Updating of Land Cover Classifications, a.k.a. BULC) which is capable of performing updates on land cover classifications by applying Bayesian statistics to classified images on a pixel-by-pixel basis. We have tested it on a study area around the Roosevelt River at the border between the Amazonas and Mato Grosso states in Brazil. As inputs into the algorithm, we used satellite imagery from Orbiew, Aster, CBERS, Landsat 1, 2, 7 and 8; classified imagery from the Finer Resolution Observation and Monitoring - Global Land Cover program and MODIS; as well as aerial imagery, at spatial resolutions ranging from 5 m to 500 m. We found that our algorithm is capable of providing up-to-date classifications despite relatively poor quality input classifications, in addition to being robust to using varied input data sources. Future developments of the BULC algorithm include testing the fidelity of its updates with that of pre-existing classification updates from the National Land Cover Database (NLCD) collection in the Las Vegas area between 1992, 2001 and 2006.