The classification of land cover is crucial due to increasing demands and population expansion. Segmentation of the terrain is utilized in environmental monitoring to accurately identify and delineate areas of agriculture, buildings, and water bodies. Satellite imagery is quite valuable for the purpose of segmentation. Deep learning models offer automated and precise segmentation of satellite imagery. The U-Net architecture is widely employed as an encoder-decoder structure for the purpose of segmenting medical and remote sensing images. This study investigates the process of dividing satellite images into segments using an optimized U-Net model. Our model utilizes improved equal weights to calculate loss functions. In addition, we have implemented a dropout layer for each convolutional block in order to decrease the computational time required for feature extraction in the encoder layer of the optimized U-Net model. Our study examines the Dubai dataset, which has six distinct classes. The optimized U-Net model we propose achieves an accuracy of 81.5% and Intersection over Union (IoU) values of 61.7% for segmenting the dataset. These metrics are acquired with a reduced number of epochs, hence decreasing the computational expense.