Multi-sensor detection of forest-cover change across 45 years in Mato Grosso, Brazil

Abstract The ongoing march toward freely available, highly pre-processed satellite imagery has given both researchers and the public unprecedented access to a vast and varied data stream teeming with potential. Among many sources, the multi-decade Landsat archive is certainly the best known, but legacy and current data from other sensors is available as well through the USGS data portals: these include CBERS, ASTER, and more. Though the particular band combinations or non-global missions have made their integration into analyses more challenging, these data, in conjunction with the entire Landsat record, are available to contribute to multi-decade surveys of land-cover change. With the goal of tracing forest change through time near the Roosevelt River in the state of Mato Grosso, Brazil, we used BULC and Google Earth Engine to fuse information from 13 space-borne imagers capturing 140 images spanning 45 years. With high accuracy, the resulting time series of classifications shows the timing and location of land-use/land-cover change—both deforestation and regrowth—at sub-annual time scales. Accuracy estimates showed that the synthesized BULC classification time series was better than nearly all of the single-day image classifications, covering the entire study area at sub-annual frequency while reducing the impact of clouds and most unwanted noise as it fused information derived from a wide array of imaging platforms. The time series improved and gradually sharpened as the density of observations increased in recent decades, when there were three or more clear, higher-resolution views of a pixel annually from any sensor combination. In addition to detailing the methodology and results of multi-source data fusion with the BULC approach, this study raises timely points about integrating information from early satellite data sources and from sensors with footprints smaller than Landsat's. There are decades of research deriving sensor-specific techniques for classifying land use and land cover from a single image in a variety of settings. The BULC approach leverages the many successes of single-sensor research and can be used as a straightforward, complementary tool for blending many good-quality mapped classifications from disparate sources into a coherent, high-quality time series.