Land-use and land-cover (LULC) classifications are important in many aspects of ecology such as conservation management, understanding species distributions, and recording changes on Earth’s surface at decadal time scales. Despite their importance, due to the expenses in time and effort to create these classifications, a published high-resolution classification can be 2-5 years old at the time of publishing, and only updated periodically. Annual classifications, while very valuable, have to date only been done at coarse (e.g., 300m, 1km) spatial resolutions. By using the new Unsupervised Bayesian Updating of Land Cover (BULC-U) algorithm, we are able to downscale a well-known LULC classification at 300-m resolution to match 30-m Landsat data while preserving accuracy and land-cover categories. BULC-U uses machine-learning techniques to segment and classify Landsat imagery to produce a classification. BULC-U then merges the coarse LULC classification with the unsupervised Landsat classification, using Bayesian statistics and confusion matrices to create a downscaled land use classification. This technique allows us to downscale land use classifications as well as representing the LULC from an earlier or later time period.By downscaling land cover classification we can make finer resolution land cover maps which are more precise and can be used to better inform research and policy decisions.
Results/Conclusions
In this study, we downscaled GlobCover 2009 in the state of Mato Grosso in Brazil. The GlobCover 2009 land cover map is a 300m resolution land cover classification created in 2011. We combined that LULC classification with Landsat 5 imagery from 2009 – 2010 in BULC-U. This resulted in an LULC classification that agreed very strongly with the GlobCover classification, but with 10 times finer spatial resolution. With its 30-m resolution, the BULC-U result was able to represent finely scaled features such as rivers and farmland as assessed with Google Earth Engine’s high-resolution satellite imagery. Furthermore, the unsupervised BULC algorithm was able to fix stray classification errors in GlobCover. If BULC-U’s success can be extended elsewhere, we may be able to refine existing LULC classifications that are reliable and frequent but coarse. If successful, this approach can be attempted across very large areas and across long time scales, which will improve our understanding of changes on Earth through time.