Thermal imaging is vital in numerous applications, such as surveillance, medical, and industrial inspections. However, they acquire images with typically lower spatial resolution than RGB cameras, resulting in limited details and effectiveness. This paper proposes an unsupervised approach for thermal image Super-Resolution (SR) referred as UnTherISR, aiming to enhance the spatial resolution of thermal images without the need for paired Low-Resolution (LR) and High-Resolution (HR) data. The proposed method leverages deep learning techniques, incorporating U-Net based CycleGAN architecture with Spatial Attention Block (SAB). Thus, by exploiting the inherent structures and textures in thermal images, the proposed approach effectively reconstructs HR thermal images, surpassing the limitations imposed by the hardware constraints of thermal cameras. The qualitative and quantitative comparisons of UnTherISR with the other existing methods demonstrate the superior performance and generalization ability for upscale factor \(\times 2\) and \(\times 4\) .