Land-use/land-cover mapping in coastal areas is a foundational yet significant pixel-wise classification work. Fully supervised semantic segmentation models have recently achieved tremendous success on this topic, but the limited annotation data will give rise to severe performance degradation. Besides, the ground categories in coastal areas characterized by great diversities may further exacerbate the adverse effects. In this paper, we propose a novel semi-supervised semantic segmentation framework based on pseudo supervision to address these issues. We first focus on the feature representation and impose perturbation to enforce consistency constraints. Unlike the naive distance measurement used in traditional consistency training, the predictions formalized with one-hot encoding from the auxiliary segmentation network are leveraged as the online pseudo supervisions. It possesses a lightweight structure and is easily generalized to multi-level architecture for exploring preferable behavior on lower-scale objects. In addition, we introduce a plain pseudo-labeling scheme to further improve the segmentation results. Its used offline pseudo supervisions are sampled from the first-stage predictions via adopting a class-wise soft version to confidence thresholds. We measure the proposed framework on two typical coastal datasets and compare it with other state-of-the-art methods. The experimental results demonstrate its excellent and competitive performance on land-use/land-cover mapping within semi-supervised scenarios.