ADAPTIVE SPATIAL SAMPLING WITH ACTIVE RANDOM FOREST FOR OBJECT-ORIENTED LANDSLIDE MAPPING

Active learning (AL) is a powerful framework to reduce labeling costs in supervised classification. However, spatial constraints on the sampling design have not yet received much attention and still pose problems for the application of AL on remote sensing data. In this study such issues are addressed in the context of landslide inventory mapping and it is shown that region-based query functions that focus the labeling efforts on compact spatial batches may provide several advantages over point-wise queries.