Multitask Active Learning for Characterization of Built Environments With Multisensor Earth Observation Data

In this paper, we propose a multitask active learning (AL) framework for an efficient characterization of buildings using features from multisensor earth observation data. Conventional AL methods establish query functions based on a preliminary trained learning machine to guide the selection of additional prior knowledge (i.e., labeled samples) for model improvement with respect to a single target variable. In contrast to that, here, we follow three multitask AL metaprotocols to select unlabeled samples from a learning set which can be considered relevant with respect to multiple target variables. In particular, multitask AL methods based on multivariable criterion, alternating selection, rank combination, as well as hybrid approaches, which internalize multiple principles from the different metaprotocols, are introduced. Thereby, the alternating selection strategies implement a so-called one-sided selection (i.e., single-task AL selection for a reference target variable with simultaneous labeling of the residual target variables) with a changing leading variable in an iterative selection process. The multivariable criterion-based methods and rank combination approaches aim to select unlabeled samples based on combined single-task selection decisions. Experimental results are obtained from two application scenarios for the city of Cologne, Germany. Thereby, the target variables to be predicted comprise building material type, building occupancy, urban typology, building type, and roof type. Comparative model accuracy evaluations underline the capability of the introduced methods to provide superior solutions with respect to one-sided selection and random sampling strategies.