Effect of inter-image spectral variation on land cover separability in heterogeneous areas

Supervised classification is a popular approach for deriving land cover data from satellite imagery, but the collection of suitable training data of large areas is expensive. Signature extension has been proposed as a method of limiting the number of training areas. Signature extension is particularly difficult in large, heterogeneous areas where the spectral characteristics of land cover classes are highly variable.The quantification of spectral separability can be used to determine to what extent a set of training areas collected in a small area can be extended to classify a larger area. This article investigates the changes in spectral separability of land cover classes in an increasing geographical area. A highly heterogeneous study area, containing nine different vegetation biomes, was chosen. Separability analyses were carried out on four Landsat-8 scenes that were sequentially mosaicked. The effect of multi-seasonal imagery on separability was also investigated. The results show that the mean spectral separability did not change when the geographical area was increased. We conclude that supervised classification with a small subset of training data should be possible in the chosen study area, since there is high separability between the classes. Some classes, however, require multi-temporal imagery as input.