Semantic segmentation of nighttime scenes poses a significant challenge in autonomous driving. While unsupervised domain adaptation offers an effective solution, existing methods disregard the relationship of knowledge transfer across different domains, which is crucial for improving model generalization ability. In this paper, we propose a Meta-Learning based Knowledge Distillation (MLKD) method tailored for adapting models trained on a source domain (daytime scene) to target domains (nighttime scenes). Firstly, we propose a brightness adjustment module based on the fast Fourier transform, which generates source images resembling the target scene in the latent domain without additional training burden. Secondly, we introduce a mask-based consistency constraint to extract knowledge from complementary latent images. This enables the model to capture rich spatial contextual relationships in scenes with lighting variations, resulting in more compact representations. Thirdly, we construct a bi-level meta-learning framework that transfers cross-domain knowledge learned from the pair of “source-to-latent” to enhance the adaptation of “latent-to-target”. Extensive experiments on benchmark datasets, i.e., Dark Zurich and ACDC, show that our MLKD achieves state-of-the-art performance, demonstrating the effectiveness of our approach in nighttime semantic segmentation.