Studying the efficacy of intelligent systems to successfully detect ice accumulation on wind turbines in cold climates has been gaining traction in recent years. In this study, both visualization and segmentation techniques were utilized in order to compare their respective results in the detection of ice on wind turbines. In pursuit of this objective, photos of wind turbines, taken under different conditions, were analyzed. To correctly classify objects automatically, pre-trained models of Resnet-50, VGG-16, VGG-19, and Inception-V3, were used. The deep learning approaches used to reliably predict the exact position of icing on wind turbine blades, including the visualization techniques Grad-CAM, Grad-CAM ++, and Score-CAM, proved to have adequate reliability with Score-CAM subsequently identified as the best visualization technique for localization. Additionally, the U-Net segmentation approach was used to delineate icing area boundaries. The U-Net approach was compared with the best visualization technique and pre-trained model to evaluate the visualization efficiency in different situations, including near and far views of a wind turbine, ice density, and light. Results showed that these methods have a high degree of accuracy in detecting ice on wind turbines.