Authors: Tolga Bakirman*, Istanbul Technical University, Elif Sertel, Istanbul Technical University
Topics: Remote Sensing
Keywords: deep learning, building segmentation, U-Net, DeepLabV3+, optical imagery
Session Type: Virtual Paper
Start / End Time: 4:40 PM / 5:55 PM
Room: Virtual 43
Presentation File: No File Uploaded
Building extraction with remote sensing plays a key role in urban applications such as temporal monitoring and LULC. It is a challenging process due to varying structures and complicated backgrounds. The rapid development in deep learning makes it an ideal tool for this task. In this study, we aim to use deep learning to segment buildings in Istanbul which have irregular urbanization with complex buildings. We have generated a private building dataset in accordance with this purpose. We have used pan-sharpened Pléiades imageries acquired in 2020 which has 3 bands (RGB), 8-bit radiometric and 50 cm spatial resolution. Currently, the created dataset covers ~50 km2 area and has around 14,000 labels. We have cropped images into 512x512 pixels and split the dataset as 70% training (395 images), 20% validation (113 images) and 10% testing (56 images). The first experiments have been carried out using U-Net architecture with DenseNet encoder. The derived metrics were 0.9623, 0,6591 and 0,7867 for accuracy, IoU and F1-Score, respectively. The further experiments were conducted with U-Net architecture with EfficientNet encoder and DeepLabV3+ architecture with SEResNeXt encoder. Obtained accuracy, IoU and F1-score metrics were 0.9551, 0.9465, 0.9718 and 0.9548, 0.9476, 0.9700 for U-Net and DeepLabV3+, respectively. The preliminary results show that the trained models are successful at segmenting buildings and can be used efficiently. However, they seem to fail extracting the boundaries of industrial buildings. Additionally, extracted boundaries are smoother than labels. Therefore, some post-processing algorithms need to be applied in order to obtain more realistic boundaries.