Land cover mapping of large areas using chain classification of neighboring Landsat satellite images

摘要: Satellite imagery is the major data source for regional to global land cover maps. However, mapping of large areas with medium-resolution costly and often constrained by lack good training validation data. Our goal was overcome these limitations, test chain classifications, i.e., classification Landsat images based on information in overlapping neighboring scenes. The basic idea classify one scene first where ground truth available, then using overlap area as We tested a forest/non-forest Carpathian Mountains horizontal six scenes, two vertical chains scenes each. collected extensive from Quickbird classifying radiometrically uncorrected Support Vector Machines (SVMs). SVMs classified 8 overall accuracies between 92.1% 98.9% (average 96.3%). Accuracy loss when automatically 1.9% average. Even resulted only an accuracy 5.1% last image compared reference independent image. Chain thus performed well, but we note that can be applied classes are well represented As long this constraint met though, powerful approach especially varying availability.