Monitoring conterminous United States (CONUS) land cover change with Web-Enabled Landsat Data (WELD)

Abstract

Egorov, A.; Potapov, P V.; Stehman, S V.; Tyukavina, A; Turubanova, S A.; Roy, D. P.; Goetz, S J.; Loveland, T R.; Ju, J; Kommareddy, A.; Kovalskyy, V.; Forsyth, C; and Bents, T, "Monitoring conterminous United States (CONUS) land cover change with Web-Enabled Landsat Data (WELD)" (2014 Forest cover loss and bare ground gain from 2006 to 2010 for the conterminous United States (CONUS) were quantified at a 30 m spatial resolution using Web-Enabled Landsat Data available from the USGS Center for Earth Resources Observation and Science (EROS) (http://landsat.usgs.gov/WELD.php). The approach related multi-temporal WELD metrics and expert-derived training data for forest cover loss and bare ground gain through a decision tree classification algorithm. Forest cover loss was reported at state and ecoregional scales, and the identification of core forests' absent of change was made and verified using LiDAR data from the GLAS (Geoscience Laser Altimetry System) instrument. Bare ground gain correlated with population change for large metropolitan statistical areas (MSAs) outside of desert or semi-desert environments. GoogleEarth™ timeseries images were used to validate the products. Mapped forest cover loss totaled 53,084 km 2 and was found to be depicted conservatively, with a user's accuracy of 78% and a producer's accuracy of 68%. Excluding errors of adjacency, user's and producer's accuracies rose to 93% and 89%, respectively. Mapped bare ground gain equaled 5974 km 2 and nearly matched the estimated area from the reference (GoogleEarth™) classification;however, user's (42%) and producer's (49%) accuracies were much less than those of the forest cover loss product. Excluding errors of adjacency, user's and producer's accuracies rose to 62% and 75%, respectively. Compared to recent 2001-2006 USGS National Land Cover Database validation data for forest loss (82% and 30% for respective user's and producer's accuracies) and urban gain (72% and 18% for respective user's and producer's accuracies), results using a single CONUS-scale model with WELD data are promising and point to the potential for nationalscale operational mapping of key land cover transitions. However, validation results highlighted limitations, some of which can be addressed by improving training data, creating a more robust image feature space, adding contemporaneous Landsat 5 data to the inputs, and modifying definition sets to account for differences in temporal and spatial observational scales. The presented land cover extent and change data are available via the official WELD website (ftp://weldftp.cr.usgs.gov/CONUS_5Y_LandCover/ftp://weldftp.cr.usgs. gov/CONUS_5Y_LandCover/).

Annual Detection of Forest Cover Loss Using Time Series Satellite Measurements of Percent Tree Cover

“…At the Landsat resolution, global VCF tree cover products have been generated for 2000 and 2005 [48]. Recent advances in remote sensing demonstrated that Landsat images could be used to create near cloud-free composites by exhaustively mining the Landsat archive [20,74], and the annually composited Landsat images could also be used to characterize land cover and change [20,75]. This may create an opportunity to produce continuous fields of land cover at annual time steps using Landsat data.…”

Landsat-Based Land Cover Change in the Beijing-Tianjin-Tangshan Urban Agglomeration in 1990, 2000 and 2010

“…Landsat images have been widely used in land cover classification because of their stable imaging quality [1,31,32,[58][59][60]. In this study, Level 1 terrain-corrected (L1T) Landsat images were selected as the primary data source for the land cover classifications.…”

Rapid Land Cover Map Updates Using Change Detection and Robust Random Forest Classifiers

“…Compositing procedures are applied independently on a per-pixel basis to the gridded WELD time series to reduce cloud and aerosol contamination, fill missing values due to the SLC failure (that removed about 22% of each Landsat ETM+ image [56]), and to reduce data volume. The WELD data have an established land cover mapping provenance and have been used, for example, to make land cover, land cover change, burned area, and field size maps at national scale [24,[57][58][59].…”

“…Multi-date composites are not suitable as input to the basic IRMAD algorithm, since only a single linear transformation is calculated per composite pair, which would not accommodate all the potential date combinations present in the composites. Although several forest monitoring systems perform change detection based on annual rank-based metrics [28,58] and it would have simplified the LALCUM system, we expected that the multiple forms of land cover change which need to be detected in our study area would require a more sensitive spectral change detection method. Since almost 22% of each of the bi-temporal Landsat ETM+ image is lost due to the SLC-off error, change detection could only be effectively applied across about 66% of the study area.…”