BUILDING FOOTPRINT SEGMENTATION USING TRANSFER LEARNING: A CASE STUDY OF THE CITY OF MELBOURNE

Abstract

Earth observation data including very high-resolution (VHR) imagery from satellites and unmanned aerial vehicles (UAVs) are the primary sources for highly accurate building footprint segmentation and extraction. However, with the increase in spatial resolution, smaller objects are prominently visible in the images, and using intelligent approaches like deep learning (DL) suffers from several problems. In this paper, we outline four prominent problems while using DL-based methods (P1, P2, P3, and P4): (P1) lack of contextual features, (P2) requirement of a large training dataset, (P3) domain-shift problem, and (P4) computational expense. In tackling P1, we modify a commonly used DL architecture called U-Net to increase the contextual feature information. Likewise, for P2 and P3, we use transfer learning to fine-tune the DL model on a smaller dataset utilising the knowledge previously gained from a larger dataset. For P4, we study the trade-off between the network's performance and computational expense with reduced training parameters and optimum learning rates. Our experiments on a case study from the City of Melbourne show that the modified U-Net is highly robust than the original U-Net and SegNet, and the dataset we develop is significantly more robust than an existing benchmark dataset. Furthermore, the overall method of fine-tuning the modified U-Net reduces the number of training parameters by 300 times and training time by 2.5 times while preserving the precision of segmentation.