This paper proposes a novel Deep Covariance Alignment (DCA) model for unsupervised domain adaptive remote sensing image segmentation. The DCA model aims to improve the generalization capability of the model by explicitly aligning category features to learn shared domain-invariant discriminative feature representations. The model consists of a Category Feature Pooling (CFP) module and a novel Covariance Regularization (CR) method. The CFP module extracts category features by combining coarse outputs and deep features. The CR method enforces intra-category features to be closer and inter-category features to be further apart by regularizing the correlation between different dimensions of the features. This approach is more robust when dealing with divergent category features of imbalanced and inconsistent distributions. The DCA model is trained using a stagewise procedure to alleviate error accumulation. Experiments on the LoveDA dataset demonstrate that the proposed DCA outperforms other state-of-the-art UDA segmentation methods in both Rural-to-Urban and Urban-to-Rural scenarios. The results show that the DCA model significantly improves the generalization capability of the model by aligning category features. The code for the DCA model is available at https://github.com/Luffy03/DCA.This paper proposes a novel Deep Covariance Alignment (DCA) model for unsupervised domain adaptive remote sensing image segmentation. The DCA model aims to improve the generalization capability of the model by explicitly aligning category features to learn shared domain-invariant discriminative feature representations. The model consists of a Category Feature Pooling (CFP) module and a novel Covariance Regularization (CR) method. The CFP module extracts category features by combining coarse outputs and deep features. The CR method enforces intra-category features to be closer and inter-category features to be further apart by regularizing the correlation between different dimensions of the features. This approach is more robust when dealing with divergent category features of imbalanced and inconsistent distributions. The DCA model is trained using a stagewise procedure to alleviate error accumulation. Experiments on the LoveDA dataset demonstrate that the proposed DCA outperforms other state-of-the-art UDA segmentation methods in both Rural-to-Urban and Urban-to-Rural scenarios. The results show that the DCA model significantly improves the generalization capability of the model by aligning category features. The code for the DCA model is available at https://github.com/Luffy03/DCA.