Satellites and AI: improving building mapping?

Have you ever thought about how complex it is to map buildings using satellite images? Scientists have made it more precise with AI. A team used U-Net, a type of neural network, to detect buildings in WorldView-3 satellite photos. But there's more: they improved results in two ways. First, they found that correcting the atmosphere before sharpening images boosts accuracy by 9.4%. Then, they created a self-supervised learning (SSL) network to train the model without labels. This led to a 15.3% improvement! The method also works on a regular PC, saving resources. A step forward for urban planning and environment.

Building segmentation using satellite imagery is crucial for disciplines like urban planning, urban climate, and environmental assessment. This study uses U-Net, a convolutional architecture, to identify buildings in WorldView-3 images. Two key innovations: 1) the optimal order between atmospheric correction (AC) and panchromatic sharpening (pan-sharpening), with AC first to increase mean IoU by 9.4%; 2) a multi-task SSL network to pre-train VGG19 on 21 unlabeled images. The two pretext tasks exploit architectural features (size, distribution, multispectral data). Results: +15.3% accuracy compared to VGG19 trained from scratch. The method is competitive against other SSL approaches, with gains of 3.5–13.7%, and optimizes computational costs, running even on consumer PCs.