<p>Dense human activities and seasonal changes make the process of identifying and monitoring land use and land cover a difficult task. Machine learning techniques and Remote Sensing technology are used to generate land cover maps as a remedy for these challenges. Nevertheless, these methods necessitate parameter tuning for supervised learning models and training samples. This research presents a novel deep learning based semantic segmentation model for the precise classification of land cover by RS images. Here, the stage of pre-processing is accomplished with a Dynamic Adaptive Spatial Range filter to enrich the visual appearance and quality of the input RS image. After pre-processing, the specific types of land covers are segmented using the Improved Twofold SwinFusion Transformer U-Net model to improve classification accuracy. Moreover, the ability of the proposed model is further improved by fine-tuning the parameters using Bobcat Brownian Motion Algorithm. The proposed deep learning-based EfficientUNet Semantic Segmentation system aims for efficient land cover classification, achieving superior performance with moderate complexity. As a result, the chief general accuracy is attained with the proposed method, with 97.51% in the land coverai dataset and 99.13% in the Mumbai dataset. The ITSFT-U-Net model’s effectiveness in classifying land cover was demonstrated by its accuracy. The outcomes are quantitatively contrasted with those of the most advanced DL models and related research.</p>