Oil spill identification from satellite imagery plays a critical role in environmental monitoring, maritime safety, and disaster response. Traditional identification methods often suffer from limited accuracy due to the presence of look-alike phenomena such as low wind areas, algal blooms, or natural films. This study presents a novel deep learning-based framework utilizing Transformer-based architectures for enhanced oil spill identification in satellite images. The proposed methodology integrates two parallel data preprocessing pipelines: RealSR for high-quality image reconstruction and GrayDN for attention-based denoising, enabling the extraction of both shallow and deep features. Multiple Transformer models, including ViT, DeiT, and Swin Transformer, are trained and evaluated to assess their performance across diverse image conditions. The dual-path preprocessing enhances image clarity and suppresses noise, significantly improving classification accuracy and robustness. Experiments conducted on a curated dataset consisting of sea surface images labeled with oil spills, ships, land, and look-alikes demonstrate that Transformer-based models outperform traditional CNNs in both precision and generalization. The GrayDN + DeiT-Base combination performs best, with an F1-score of 0.932. This approach shows promise for future maritime environmental monitoring systems and offers a scalable solution for oil spill identification.