Defective road identification is a challenging and necessary task, specifically for self-driving vehicles. Self-driving cars are used to transport different objects, including human beings. These vehicles offer an easy way to transport everything within their capacity and can do so without the assistance of a human driver. In the case of self-driving cars, it is necessary to identify the defects in the roads initially. Identification of road conditions is a complex task due to potential adverse weather conditions like fog, haze, and rain. In this paper, we implement CycleGAN to generate synthetic road images including the effects of bad weather, such as rain, fog, and haze. Additionally, we apply a saliency map to draw visual attention to the damaged pixels in the very poor roadway images. We have transferred the domain knowledge by adapting their information. We have implemented deep neural networks like Vgg-16, ResNet-50, small CNN, InceptionNet v3, and Vision Transformer pre-trained on the ImageNet 21k dataset. Among all the above-mentioned deep networks, Vision Transformer outperforms others.