Automated identification of temperate conifer forests in Landsat imagery: generalization in time and space

Monitoring change in temperate conifer forests requires the combination of spatial and spectral resolutions, and temporal coverage available from the Landsat Program. One key step in the development of methods for automated monitoring of change in temperate conifer forests is an algorithm for identification of conifer forests in previously unseen Landsat images. This problem is approached through analysis of field samples collected from over 13 different Landsat Thematic Mapper (TM) scenes, and multiple dates of imagery for each scene. The central question concerns how well conifer forest can be classified when training data do not come from the image being classified. The authors´ approach to this question is to evaluate classification accuracies as training and testing data become increasingly more separated in time and space. As this separation increases, the variability between images due to factors unrelated to land cover spectral characteristics (such as sensor calibration, atmospheric conditions, and topographic variability) are expected to become more pronounced and cause confusion in the classifications. Therefore, the effects of data calibration on classification accuracies across time and space is also evaluated