Alternative Approaches for Coarse Resolution Shrub Density Mapping in Arctic Tundra: Comparing Landsat Multi-spectral, MISR Multi-spectral/multi-angular and MODIS Multi- temporal Mapping Approaches

Increases in shrub cover have been documented across arctic Alaska and have been linked to a warming climate. This study explores the utility of several remote sensing platforms for mapping shrub cover in the arctic. A 90 km x 12 km swath of IKONOS imagery acquired on a single overpass in July 2001 and in situ measurements of shrub density acquired during the summer of 2008 were used to produce a map of shrub density with a spatial resolution of 4 m/pixel. Cross validation indicated this map had a root mean squared error (RMSE) of 2.9% and a coefficient of determination (R2) of 0.88. The 4 m/pixel image was then degraded to a spatial resolution of 275 m/pixel and used as the training and validation dataset for regression tree models based on input from coarser resolution sensors. Preliminary results indicate that while a model using the six spectral bands available from a July 2000 Landsat ETM+ image reproduces the high resolution shrub density results most accurately at the 275 m spatial scale (R2 = 0.66, RMSE = 3.04), a model based on four nadir-view spectral bands and red band data from nine separate viewing angles from the MISR sensor from late June 2001 produced results only slightly less accurate (R2 =0.58, RMSE = 3.10). Using the combination of multi-spectral and multi-angular data from MISR appears to boost accuracy considerably; results from models using only the four spectral bands from the nadir view camera (R2 = 0.52, RMSE = 3.30) or only the red spectral band from all nine cameras (R2 = 0.48, RMSE = 3.5) were less accurate. A model based on a sequence of Enhanced Vegetation Index (EVI) data from MODIS from summer 2001 and 2002 produced far less accurate results (R2 = 0.25, RMSE = 4.3), although adding the EVI sequence to the MISR multi-spectral/multi-angular model did result in a very slight improvement in accuracy (R2 = 0.59, RMSE = 3.10). While Landsat has the advantage of higher spatial resolution, experience with large scale land cover mapping projects that span multiple Landsat scenes in areas where the short snow-free season coincides with persistent cloud cover has demonstrated that it is often impossible to find adjacent Landsat scenes with similar vegetation phenology. These preliminary results indicate that for large scale shrub density mapping where high spatial resolution is not required, using a combination of multi- spectral, multi-angular, and multi-temporal data may be a viable alternative to using Landsat or similar medium resolution sensors.