High-resolution urban land-cover classification using a competitive multi-scale object-based approach

Abstract

In this study, a two-step classification procedure was used for classifying urban land cover. First, a hierarchy of seven image segmentations of different scales was created for an urban scene, and preliminary classifications were performed for each of the segmentations using a classification algorithm that provides the probability that a segment belongs to a land-cover class in addition to the class assignment. A higher probability for the assigned class indicates that a segment is more likely to have been classified correctly. The class assignment and probability for image segments in each of the six coarser segmentations were added to the segments that they contained in the finest scale segmentation to allow for a comparison across all of the scales. For the final classification, the finest scale segments were assigned to the class that gave the highest probability in any of the segmentations. The incorporation of preliminary class assignments and probabilities across multiple scales led to an increase in overall accuracy from 78.1% to 82.1%.