With the rapid growth of remote sensing applications in various fields, the need for accurate semantic segmentation of remote sensing images has become increasingly urgent. However, the difficulty in acquiring high-quality semantically labeled data has limited the effectiveness of traditional semantic segmentation models that rely on labeled data in remote sensing. This has resulted in a need for alternative approaches to improve the accuracy of remote sensing data analysis. We presents a semi-supervised semantic segmentation framework based on consistency learning. On the one hand, a unified perturbation strategy that unifies the image-level perturbation and the feature-level perturbation into separate streams to exploit a larger perturbation space. On the other hand, a dual-stream perturbation strategy is designed to fully explore the predefined image-level perturbation space and take advantage of comparative learning for discriminative characterization. The framework integrates the above two strategies and shows superior performance for semantic segmentation tasks in the case of remote sensing images with sparsely labeled data.