Building Footprint Extraction Using Deep Learning Semantic Segmentation Techniques: Experiments and Results

Abstract

Deep Learning Semantic Segmentation are techniques based on convolutional neural networks that have been employed in several research applications in Computer Sciences. In recent years, Remote Sensing researchers have used such techniques and achieved remarkable results. In this research paper we employ semantic segmentation techniques to extract building footprints from remote sensing imagery. We use a custom dataset called Brazilian Army Geographic Service Building Dataset to train several neural network architectures such as U-Net and FPN, combined with the following backbones: EfficientNet-B0, EfficientNet-B1, SE-ResNeXt-101 and ResNet-152. To train the mentioned structures, a framework based on Keras and Tensorflow called segmentation_models_trainer (https://github.com/phborba/segmentation_models_trainer) was used. Transfer Learning from ImageNet weights were used, as well as data augmentation on training images.