This document proposes a method for automatically detecting compound structures from multiple hierarchical segmentations of remote sensing images. Compound structures contain spatial arrangements of primitive objects like buildings, trees, and roads. The method models compound structures as probabilistic region processes and learns their appearance and spatial models from training data. Candidate regions are extracted from hierarchical segmentations, and a constrained region selection framework is used to detect compound structure instances by selecting coherent subsets of regions that satisfy constraints. Approximate inference is performed using Markov chain Monte Carlo sampling or quadratic programming under constraints.