Efficient building segmentation with a lightweight U-Net in high-resolution aerial imagery

The precise and accurate semantic segmentation of complex scenes from remote sensing images has always been a challenging task. Thanks to advances in deep neural networks, segmentation models, especially U-Net and its variants, have shown promising segmentation results. However, the enhanced performance of the currently popular deep segmentation models comes at the cost of time and computational efficiency. Moreover, these models often struggle to deliver accurate segmentation results when computational resources are constrained, as is common in large-scale or operational remote sensing workflows. Therefore, this study presents a significantly small, fast, and computationally efficient version of a U-Net, called TinyEffUnet. To verify the effectiveness of the proposed lightweight model, we conduct extensive experiments on a building extraction dataset – INRIA aerial scene segmentation dataset. Compared to a canonical U-Net, TinyEffUnet achieves superior intersection over union (IoU) performance in extracting building structures from aerial imagery with only 1.8M parameters and 5.08B floating point operations (FLOPs), demonstrating its efficiency and suitability for large-scale remote sensing applications. This efficiency also makes it well-suited for deployment on satellite payloads with computational constraints.