A Building Footprint Extraction from UAV Imagery Using Deep Learning.

Abstract

Building footprint extraction from UAV (Unmanned Aerial Vehicle) imagery data has been an active research topic in the photogrammetric and remote sensing communities in the past two decades. The traditional methods for building extraction from high-resolution imagery data are time-consuming and may not provide desired results. Recently, effective high-level approaches have been developed for buildings footprint extraction. However, their efficiency must be balanced by reducing the processing time required to obtain acceptable results using high-resolution imagery. This paper introduces an automatic method to extract building footprints from UAV imagery using deep learning algorithms. A Mask R-CNN (Region Based Convolutional Neural Network) model has been applied to building footprint extraction. For the building extraction, three experiments have been achieved. For the algorithm testing, two study areas have been selected. An orthophoto has been produced for each study area using photogrammetric software based on UAV imagery. Three experiments have been achieved for building extraction from the study area. The first experiment was based on using the pre-trained model only. In the second trial, the raw model was trained based on the study area only was used. While in the third trial, the model was trained based on fine-tuning (based on satellite imagery) and the pre-trained model with UAV training data. The analysis showed that the highest accuracy rate of the building footprint extraction increased to around 95% through using fine-tuning and more data sets in the model training, specifically with similar data sets to the study area.

Keywords

• building footprint
• deep learning
• mask r-cnn
• uav
• fine-tune