Loss Function Analysis for Building Extraction from Remote Sensing Images

Building extraction from remote sensing images is a difficult task with significant implications for environmental monitoring and human activity area construction. Deep learning (DL) methods, particularly convolutional neural network (CNN), have emerged as promising approaches for building extraction in remote sensing. Despite their effectiveness in feature extraction, CNN-based methods often suffer from information loss, leading to performance degradation. To address this issue, researchers have proposed various loss functions to enhance building extraction performance. In this paper, we investigate the impact of various loss functions on the semantic segmentation model (UNet) for building extraction. Specifically, we evaluate five commonly used loss functions, including cross-entropy loss, focal loss, Tversky loss, dice loss, and Jaccard loss. Our experiments are conducted on two widely available satellite datasets, namely, WHU-I and WHU-II. The results demonstrate that Jaccard loss outperforms the other loss functions for both datasets, highlighting its effectiveness in building extraction tasks.