Generalizability in convolutional neural networks for various types of building scene recognition in High-Resolution imagery

Abstract

Building recognition is a core task for urban image classification (mapping), especially in optical high-resolution imagery. Convolutional Neural Networks (CNNs) have recently achieved unprecedented performance in the automatic recognition of objects (e.g. buildings, roads, or trees) in high-resolution imagery. Although these results are promising, questions remain about generalizability. This is a great challenge, as there is a wide variability in the visual characteristics of the building image scene across different geographic locations. CNNs are overfitted with limited and low diversity samples and are tested on the same or nearby geographic locations. In this work, we propose two scenarios with regard to transfer learning CNN features for building scene classification. We also investigate the generalizability of CNNs for building recognition across different geographic locations. The results of the two scenarios show that the final model, generalizable in different geographic locations, unseen areas.