Performance evaluation of shallow and deep CNN architectures on building segmentation from high-resolution images

Abstract

Building extraction from high-resolution images has been studied extensively for its great importance in obtaining geographical information. As an advanced machine learning technique, deep learning has achieved great progress along with developments in hardware and larger datasets. In this study, the performance evaluation of convolutional neural network architectures in building segmentation from high-resolution images was investigated. Four U-Net based architectures were generated and their performances were compared with each other and to the U-Net. Models were trained and tested on datasets that were prepared using the Inria Aerial Image Labelling Dataset and the Massachusetts Buildings Dataset. On the INRIA test dataset, Deeper 1 architecture provided 0.79 F1 and 0.66 IoU scores. Deeper 1 was followed by Deeper 2 and U-Net architectures, both with an F1 score of 0.78 and an IoU score of 0.65. On the Massachusetts test dataset, the U-Net architecture provided 0.79 F1 and 0.66 IoU scores. This architecture was followed by Deeper 2 with 0.78 F1 score and 0.65 IoU score, and Shallower 1 and Deeper 1 architectures both with 0.77 F1 score and 0.64 IoU score. The successful results of Deeper 1 and Deeper 2 architectures show that deeper architectures can provide better results even if there is not too much data. Also, Shallower 1 architecture appears to have a performance not far behind deep architectures, with less computational cost, and this shows usefulness for geographic applications.