It is our goal to address the challenge of fusing satellite images. The currently available pansharpening models fail to strike a balance between clear boundaries and rich spatial–spectral details. We propose an unsupervised remote sensing image pansharpening model based on a triple-discriminator generative adversarial network (TDP-GAN) that combines low-resolution multispectral (LRMS), panchromatic (PAN), and structural gradient (SG) images. This is the first attempt to produce pansharpened images using triple-discriminator adversarial networks. TDP-GAN trains directly on full-resolution images without downsampling and achieves the balance. First, the SG image is generated by superposition of the structural information in the horizontal and vertical directions from the PAN image using the first-order forward differential operator. Then, we input SG images along with PAN and LRMS images into the multistream generator. Finally, a special nonreference loss function is designed to improve image generation effects by reducing the loss of spatial structural information and spectral detail in generated images. The proposed model can optimize training processes, accelerate network convergence, and deepen the gradient information exchange between the generator and discriminators. Experiments conducted on publicly available datasets show that TDP-GAN performs better than state-of-the-art methods on both quantitative and qualitative metrics.