In recent years, convolutional neural networks (CNNs) have excelled in remote sensing image super-resolution reconstruction (RSISR) tasks, becoming the predominant algorithms in this domain. However, these models primarily leverage the dependency of high-resolution (HR) images on low-resolution (LR) counterparts during the super-resolution (SR) forward process, neglecting mutual dependencies. To address the ill-posed nature of one-to-many mappings and enhance reconstruction performance, this paper proposes Adaptive Enhanced Reversible Flow Model (AERNet), an image SR algorithm based on invertible neural networks. AERNet treats image degradation and reconstruction as invertible transformations, where LR and HR images mutually project into each other's spaces. This mutual dependency optimizes distribution mapping across LR and HR images, constraining the solution space effectively in both forward and inverse directions. Integrating a multi-path adaptive feature fusion group and a global interaction enhancement module enhances the network's adaptability, improving its capability to fuse and enhance feature information. This approach enables more accurate processing of key image details and regions. Experimental results demonstrate AERNet's superior performance on two benchmark remote sensing datasets.