Buildings' Classification Using Very High-Resolution Satellite Imagery

Buildings' classification using satellite images is gaining importance for several applications, such as damage assessment, resource allocation, and population estimation. This work focuses on two specific classification tasks: building-type classification (residential or non-residential) and building-damage assessment (damaged or not-damaged). These two tasks have become of great interest recently, especially considering current global conflicts and natural disasters, such as the Ukrainian conflict, the 2023 Turkey–Syria earthquake, and the recent tragic flood in Libya's Derna.

Existing building-type classification approaches combine optical satellite images with LIDAR data or street view images. We propose here to rely solely on RGB satellite images and follow a 2-stage deep learning-based approach, where buildings' footprints are extracted using a semantic segmentation model, followed by the classification of the segmented images. To the best of our knowledge, this is the first Building-Type Classification (BTC) method solely based on RGB aerial imagery. Optical imagery is cost-efficient, widely available, and can be tasked and acquired quickly following a military conflict or natural disaster.

Supervised deep learning algorithms rely on high-quality labeled datasets. Four Building-Damage Assessment (BDA) datasets are considered, analyzed, and compared to choose the appropriate one for our research. For the scope of this work, we define a damaged building as a partly or wholly demolished building that might result from armed conflicts, earthquakes, tornadoes, and hurricanes. On the other hand, due to the need for an appropriate residential/non-residential building-type classification dataset, we introduce a new high-resolution satellite imagery dataset called the Beirut Buildings Type Classification (BBTC) dataset.

We conduct experiments to select the best hyper-parameters and model architecture and propose an extra transfer learning stage that outperforms the classical method. Finally, we validate the performance of the proposed schemes, where BTC achieved a 94.8% accuracy using RexNet as a backbone, and BDA scored 97.3% accuracy using Focal loss and Adam optimizer. With the extra transfer learning stage, BDA improved in the last percentages of the accuracy and F1-score to reach 98.96% and 99.4%, respectively.