<p>In the current era, object detection from satellite images has become essential in various applications, from environmental surveillance to urban development. U-Net, one of the popular encoder-decoder architectures, has proven to be a strong performer in segmentation tasks. Different U-Net variants, such as the baseline U-Net, Attention U-Net, and Attention Res-UNet (Attention Deep Residual U-Net), offer architectural innovations that yield higher segmentation accuracy in specific contexts. This research presents an ensemble approach that combines the benefits of various U-Net models to refine overall segmentation quality. U-Net variants are comprehensively explored for segmentation on satellite images, emphasising how they perform in the semantic segmentation of an aerial images dataset. Interestingly, the ensemble model outperforms single models considering IoU and dice score, delivering a staggering IoU of 93.2 percent, 90 percent pixel accuracy, and a dice score of 94.1 percent, accentuating its outstanding performance. The findings indicate the potential of the suggested ensemble method, that integrates complementary U-Net architectures to deliver state-of-the-art results in satellite image analysis.</p>