Building detection in multi-source aerial data with imbalanced training samples: an approach based on the Bayesian Vector Quantizer

ABSTRACT

The problem of building detection in multi-source aerial data has a large variety of applications from map updating to the detection of illegal construction. The development of a capability to integrate multispectral and LiDAR technologies has been proved to be the most effective strategy in dealing with the problem. An automated and combined approach enables such joint capabilities to process data, reducing human effort that is usually limited to the creation of training sets. This aspect plays a key role in getting accurate results and it is central to dealing with the problem of an imbalanced data set. We tailor the Bayesian Vector Quantizer algorithm (BVQ) to the problem of building detection from multi-source high-resolution aerial data like LiDAR with a focus on the imbalanced data set problem. The result is a methodology that is optimised towards the solution of strongly imbalanced problems where noise is present and where the number of training samples for buildings over classes like trees, land and grass is one order of magnitude lower. The formulation is compared with other well-adopted approaches which highlight the relative strengths of the BVQ approach.

KEYWORDS:

• BVQ
• imbalanced data classification
• building detection
• high-resolution multispectral and LiDAR imagery

Acknowledgement

The authors would like to thank Carla Nardinocchi at the University of Rome La Sapienza and Kourosh Khoshelham at the University of Melbourne for having provided the datasets and domain knowledge as well as Jack Dyson at the Università Politecnica delle Marche in Ancona for having reviewed and provided suggestions to improve the paper.