Marine pollution introduces harmful substances into the ocean, affecting ecosystems, marine life, coastal communities, and the global economy. Classifying these pollutants is essential for identifying their sources and assessing their ecological impact. Computer vision techniques are used to automate analysis and enhance the accuracy of detecting and classifying marine pollutants as visual identification results in underreporting of pollutants. Sentinel-2 Multispectral images have very low visibility of pollutants. The proposed method uses (i)High quality Sentinel-2 multispectral thermal images generated by Stable Diffusion Thermal Image Generator highlights temperature variations for better classification (ii) Transverse Dyadic Wavelet Transform (TxDyWT) to pre-process the thermal images as it retains structural details for classifying pollutants.(iii) Denoising Convolutional Neural Network (DnCNN) optimized with Hippopotamus Optimization Algorithm enhances images and Vision transformer (ViT) is employed to classify as microplastics, sediments and oil spills by identifying subtle patterns in pollutants. The proposed methodology identifies fragments of microplastics as small as 0.5 mm, large-scale oil spills, and hydrogenous sediments. The detection accuracy for microplastics, oil spills, and sediments is approximately 95%.