Semantic segmentation is made more difficult by the high intraclass variability, spectral heterogeneity, and thin object boundaries found in high-resolution satellite imagery. To overcome such hurdles, one needs to use architectural paradigms that can understand both the overall sketch and the finer details of a given location. A cutting-edge deep learning platform, DeepSegNet, is proposed in this study. Its single purpose is semantic pixel-by-pixel satellite data segmentation. The architecture combines three important innovations in a hierarchical encoder-decoder network: an adaptive receptive field learning module, an Attention-Enhanced Skip Connections (AESC) feature propagation module, and an Adaptive Residual Refinement Module (ARRM) structural correction module. With uniform preprocessing and augmentation techniques, the system is evaluated on two benchmark datasets, SpaceNet v2 and ISPRS Vaihingen. DeepSegNet reliably outperforms state-of-the-art techniques in quantitative findings. An efficient ablation study has confirmed all of the architectural elements. Retaining object boundaries is made more efficient, and performance on low-frequency and small-class instances is improved, as supported by qualitative studies.