Remote sensing image segmentation by active queries

Active learning deals with developing methods that select examples that may express data characteristics in a compact way. For remote sensing image segmentation, the selected samples are the most informative pixels in the image so that classifiers trained with reduced active datasets become faster and more robust. Strategies for intelligent sampling have been proposed with model-based heuristics aiming at the search of the most informative pixels to optimize model's performance. Unlike standard methods that concentrate on model optimization, here we propose a method inspired in the cluster assumption that holds in most of the remote sensing data. Starting from a complete hierarchical description of the data, the proposed strategy aims at sampling and labeling pixels in order to discover the data partitioning that best matches with the user's expected classes. Thus, the method combines active supervised and unsupervised clustering with a smart prune-and-label strategy. The proposed method is successfully evaluated in two challenging remote sensing scenarios: hyperspectral and very high spatial resolution (VHR) multispectral images segmentation.

► Active learning is proposed to update unsupervised remote sensing image segmentation.
► Novel query strategies are proposed, both in terms of nodes and segments selected.
► Spatial regularization is studied and increases the efficiency of the methods.
► The approach outperforms random and state of art active sampling in three problems.
► Visual inspection of the classification maps show the efficiency of the method.