HRI: Harmonizing Rotation Invariance for Remote Sensing Pretraining Model

In remote sensing, self-supervised pretraining methods like masked image modeling and contrastive learning have demonstrated promising image understanding capabilities. However, existing approaches largely overlook the critical challenge of rotation diversity, failing to learn robust rotation-invariant representations. To address this gap, particularly given the arbitrary orientations of ground targets and imaging angles, we propose a novel pretraining framework HRI that systematically integrates rotation-aware tasks, enabling the model to learn rotation-invariant features from multiple levels while preserving semantic consistency across rotational transformations. Specifically, we introduce a masked multi-angle reconstruction strategy that, for the first time, reconstructs masked images at various rotated angles to capture pixel-level rotation-invariant and textural features. Additionally, we exploit the rotation-averaged features for contrastive learning, which helps suppress noise and enhances the learning of robust, global rotation-invariant representations. Extensive experiments across a range of downstream tasks demonstrate that our method outperforms existing approaches in robustness and accuracy under varying rotation conditions. The code will be released at https://github.com/mint0126/HRI.