Automatic Building Extraction from VHR Remote Sensing Images Using Geoai Methods

Building footprint extraction is a crucial task in remote sensing that helps acquire accurate building information for various applications such as city planning, population estimation, and disaster management. In this study, we explored the performance of Unet, Unet++, and DeepLabV3+ segmentation architectures on a very high-resolution Wuhan University Aerial Building dataset. We used InceptionResNetV2 and SE-ResNeXt101 encoders with these segmentation models after conducting pre-experiments with multiple encoder and hyper-parameter combinations. Furthermore, we implemented transfer learning by using the shared weights of a previous building detection study. We converted the raster outputs of deep learning models to vector format to enable a better spatial comparison among different models. All models were trained on the Kaggle platform, utilizing a Tesla P100-PCIe-16GB GPU and the PyTorch library. The F-1 scores for the test dataset range between 0.9867 and 0.9897 for different experiments. As a final assessment, we visually compared our experiment results with the Segment Anything Model.