Unsupervised Domain Adaptation (UDA) is the process of learning knowledge from a labelled source domain to an unlabelled target domain, particularly in the context of remote sensing scene classification. The primary challenge in this process is the substantial cost associated with labelling and the significant discrepancies between domains. However, existing UDA methods degrade under severe domain shift and scene diversity, yielding noisy pseudo-labels and unstable target structure discovery. To address these issues, a novel UDA framework is proposed. The main focus of the framework is to develop a mapping using clustering-based pseudo-labelling that can provide a reliable and interpretable pseudo-labels to the target dataset. A deep learning-based Pareto font-driven feature-selection module is also added to fine-tune the source and target features, thereby significantly improving the performance of the scene classification model. An adaptive density-based clustering method with a two-step neural network in the clustering module is utilized to determine whether adjacent clusters should be merged, thereby maintaining clear class boundaries. To reduce the pseudo-label noise, an uncertainty-aware soft pseudo-labelling approach is implemented, based on a dynamic confidence threshold. The framework is evaluated on four remote-sensing datasets namely AID (A), NWPU (N), RSSCN7 (R), and UC Merced (U) across various domain-adaptation tasks (A → R, R → A, A → U, U → A, R → U, U → R, N → R, and R → N). The proposed approach achieves accuracy improvements of 5.80%, 1.8%, 2.79%, 7.56%, 3.50%, 7.39%, 5.35%, and 3.12% over some of the baseline methods. These results show the superiority of the proposed approach in managing domain shifts, reducing pseudo-label noise, and improving target recognition without the need for labelled target data. The source code is available at https://github.com/BinuJoseA/UDA .