Deep Learning for Building Footprint Extraction in Aerial Imagery
The Vector Institute for Artificial Intelligence published a technical blog post detailing a comparative study on deep learning methods for extracting building footprints from high-resolution satellite imagery. This research addresses the growing availability of public and private satellite data, which supports critical downstream applications like humanitarian assistance and disaster response. The study focuses on semantic segmentation, a computer vision task that labels each pixel to identify objects, specifically distinguishing buildings from background. Using the SpaceNet Building Detection V2 dataset, which includes over 300,000 building labels across cities like Las Vegas, Paris, Shanghai, and Khartoum, the team benchmarked six different architectural approaches. These included U-Net, U-Net++, Fully Convolutional Networks (FCN), and DeepLabv3, with varying backbones such as Resnet-50 and Resnet-100. The article explains how encoder-decoder architectures and skip connections facilitate precise localization and context capture. By evaluating these models, the Vector AI Engineering team aims to enhance the accuracy and robustness of geospatial analysis tools. The work highlights the transition of successful semantic segmentation techniques to specific geospatial tasks, providing insights for engineers and researchers working in applied machine learning and computer vision.
Wire timeline
Deep Learning for Building Footprint Extraction in Aerial Imagery
The Vector Institute for Artificial Intelligence published a technical blog post detailing a comparative study on deep learning methods for extracting building footprints from high-resolution satellite imagery. This research addresses the growing availability of public and private satellite data, which supports critical downstream applications like humanitarian assistance and disaster response. The study focuses on semantic segmentation, a computer vision task that labels each pixel to identify objects, specifically distinguishing buildings from background. Using the SpaceNet Building Detection V2 dataset, which includes over 300,000 building labels across cities like Las Vegas, Paris, Shanghai, and Khartoum, the team benchmarked six different architectural approaches. These included U-Net, U-Net++, Fully Convolutional Networks (FCN), and DeepLabv3, with varying backbones such as Resnet-50 and Resnet-100. The article explains how encoder-decoder architectures and skip connections facilitate precise localization and context capture. By evaluating these models, the Vector AI Engineering team aims to enhance the accuracy and robustness of geospatial analysis tools. The work highlights the transition of successful semantic segmentation techniques to specific geospatial tasks, providing insights for engineers and researchers working in applied machine learning and computer vision.
Vector Institute for Artificial Intelligence