Satellite imagery plays a vital role in a wide range of remote sensing applications, yet it often suffers from various forms of degradation including sensor noise, atmospheric disturbances, and limited spatial resolution. These degradations hinder accurate interpretation and analysis, particularly in tasks that require high spatial and spectral fidelity. In this work, we present a two-stage deep learning framework that addresses these challenges by denoising corrupted satellite images and enhancing their resolution through super-resolution techniques. For the denoising stage, we employ a Convolutional AutoEncoder, which enables the model to learn noise statistics without requiring clean target images. This is particularly beneficial in satellite imaging where ground truth is often unavailable or costly to obtain. Following denoising, we apply a super-resolution model that upscales the image while preserving fine-grained details. The integration of these two modules allows for effective removal of high-frequency noise and reconstruction of high-resolution features in a sequential and complementary manner. We evaluate our approach qualitatively on a curated set of noisy satellite images, demonstrating clear improvements in visual clarity, edge preservation, and color realism compared to the original inputs. Quantitative results further support these improvements: our model achieved a PSNR of 29.19dB and SSIM of 0.796 for denoising, and a PSNR of 32.41dB and SSIM of 0.900 for super-resolution on the EuroSAT dataset. The proposed system highlights the potential of deep learning in overcoming the limitations of conventional satellite imagery and paves the way for more reliable and interpretable Earth observation data. Future work will focus on expanding this framework to multispectral inputs, integrating end-to-end architectures, and enhancing robustness under real-world noise conditions.