Probabilistic Breaking Tie: An Active Learning Strategy To Leverage Class Hierarchy For Impervious Surfaces Classification

Urban sprawl has significantly changed the land cover of metropolis and periurban areas. In that regard, the monitoring of impervious surfaces has become crucial to mitigate their impacts on the hydrology of the watersheds and on urban micro-climates. At this end, airborne hyperspectral imaging is very well suited to map the diversity of impervious surfaces over urban areas, both in terms of scale and of discriminative power. Yet, the diversity of soil materials as well as the similarity between some artificial porous and impervious classes, is such that active learning methods are imperative to build optimal ground truths for machine learning algorithms. Besides, in the context of impervious surfaces classification, not all classes are of equal interest. In the present paper, we introduce Probabilistic Breaking Tie, a query system, built on the state-of-the-art Breaking Tie [1] heuristic, that leverages class hierarchy as a priori semantic knowledge. Our numerical experiments on the Houston University data set [2] show that our method significantly increases the pace at which active learning improves the accuracy metrics over impervious and porous classes. Code is available at https://github.com/Romain3Ch216/AL4EO.