 Isosceles is a method for hierarchical sampling of large image sets that uses affinity propagation to automate the selection of highly representative training images, reducing human bias and oversampling uninformative areas while eliminating under-sampling informative ones. It obtains a training set that reflects both between scene and within-scene variability for country-scale building extraction, resulting in an increase in accuracy of 2.2 to 4.2% compared to stratified random sampling using three, distinct model architectures. This article was authored by Benjamin Swan, Melanie Lavardier, H. Lexie Young, and others.