Semantic Segmentation of Objects from Airborne Imagery

Extraction of objects from images acquired by airborne sensors is the one of the most important topics in Aerial Photograph Interpretation (API). The task is challenging due to the very heterogeneous appearance of man-made and natural objects on the ground. Meanwhile images acquired by airborne sensors are very high-resolution, which requires high computational costs. This paper presents an efficient approach for automated extraction of objects at pixel level. We propose to combine a powerful classifier and an efficient contextual model for semantic segmentation of objects in images. Multiple image features are used to train the classifier, other features are used to learn the contextual model. We employ Random forest (RF) as classifier which allows one to learn very fast on big data. The outputs given by RF are then combined with a fully connected conditional random field (CRF) model for improving classification performance. Experiments have been conducted on a challenging aerial image dataset from a recent ISPRS Semantic Labeling Contest. We obtained state-of-the-art performance with a reasonable computational demand.