Land cover change detection: A multi-scale approach to comparing remote sensing maps

Mapping the changes in the distribution of land cover provides critical data for monitoring broad-scale ecosystem change and investigating both the effects and determinants of such change. Because of variation in classification methodology, image quality and remote sensing technology, characteristics of land cover maps based on remote sensing vary considerably. Differences in land cover maps between time periods due to change on the ground are confounded with differences attributable to methods and sensors; simply overlaying maps from two points in time may not produce an accurate spatial representation of change. In this study I used the Landscape Scripting Language (LSL) multi-scale spatial analysis package to assess the nature and magnitude of map differences for two pairs of contemporaneous forest cover maps (ca. 1991 and ca. 2002) interpreted from satellite imagery. Within-period comparisons indicate the magnitude of method-based variation, while between-period differences include both methodological and cover differences. I used LSL and its regression facility to compare maps at spatial grains ranging from 1/7 ha through 1500 ha. This analytical approach produced estimates of classification variability and identified an appropriate resolution for mapping. Regression parameters and statistics provided estimates of bias and precision; for maps from differing times, residuals were used to map forest cover change. The analysis indicated the optimal grain for mapping forest cover change from these maps was approximately100 ha. In the context of ecosystem restoration, land cover trends will identify landscapes under recent disturbance pressure and may be used as performance indicators for broad-scale evaluation of habitat restoration policies.