This paper presents a simple method for unsupervised domain adaptation (UDA) in semantic segmentation. The method, called Fourier Domain Adaptation (FDA), reduces the discrepancy between source and target distributions by swapping the low-frequency spectrum of one with the other. It is applied to semantic segmentation, where densely annotated images are available in one domain (e.g., synthetic data) but not in another (e.g., real images). Current state-of-the-art methods are complex and often require adversarial optimization to make neural networks invariant to domain selection. In contrast, FDA achieves state-of-the-art performance without any training beyond the primary task of semantic segmentation. It uses a simple Fourier Transform and its inverse to align the low-level statistics between the source and target distributions. The method is tested on two synthetic-to-real domain adaptation tasks: GTA5→CityScapes and SYNTHIA→CityScapes. Results show that FDA outperforms existing methods, including adversarial training approaches, and demonstrates that simple procedures can effectively handle nuisance variability in data. The method is robust to the choice of the spectral neighborhood size (β) and achieves significant improvements in semantic segmentation performance. FDA is also combined with multi-band transfer (MBT) to further improve performance by averaging predictions from different models trained with different β values. The method is efficient, computationally simple, and does not require complex adversarial training. It is shown to be effective in reducing domain gap and improving performance in semantic segmentation tasks.This paper presents a simple method for unsupervised domain adaptation (UDA) in semantic segmentation. The method, called Fourier Domain Adaptation (FDA), reduces the discrepancy between source and target distributions by swapping the low-frequency spectrum of one with the other. It is applied to semantic segmentation, where densely annotated images are available in one domain (e.g., synthetic data) but not in another (e.g., real images). Current state-of-the-art methods are complex and often require adversarial optimization to make neural networks invariant to domain selection. In contrast, FDA achieves state-of-the-art performance without any training beyond the primary task of semantic segmentation. It uses a simple Fourier Transform and its inverse to align the low-level statistics between the source and target distributions. The method is tested on two synthetic-to-real domain adaptation tasks: GTA5→CityScapes and SYNTHIA→CityScapes. Results show that FDA outperforms existing methods, including adversarial training approaches, and demonstrates that simple procedures can effectively handle nuisance variability in data. The method is robust to the choice of the spectral neighborhood size (β) and achieves significant improvements in semantic segmentation performance. FDA is also combined with multi-band transfer (MBT) to further improve performance by averaging predictions from different models trained with different β values. The method is efficient, computationally simple, and does not require complex adversarial training. It is shown to be effective in reducing domain gap and improving performance in semantic segmentation tasks.