<p>Remote sensing images have wide application prospects in transportation, agriculture and environmental monitoring. However, different regions of remote sensing images have the characteristics of different size, which tend to be affected by noise, leading to artifacts, thus, accurately recovering terrain texture details is challenging. This paper proposes a super-resolution reconstruction method of remote sensing images, SwinDSR, based on dimension permutation and asymmetric feature fusion. A parallel architecture combining self-attention mechanism and depthwise separable convolution is employed to achieve global and local feature fusion. To capture broader contextual information, the single-scale convolution in the reconstruction network is improved to multi-scale asymmetric convolutions, enhancing the network’s adaptability to features at different scales. Additionally, a random shuffle module is introduced, along with Sinc filters to truncate high-frequency components in the image, simulating ringing and overshoot artifacts. This constructs a multi-factor degradation network that effectively suppresses artifact generation during remote sensing image reconstruction. In addition to the reconstruction of high-quality remote sensing images, it is potential to achieve the lightweight of SwinDSR model.</p>