This paper proposes a novel remote sensing pre-training framework called Generic Knowledge Boosted Remote Sensing Pretraining (GeRSP) to learn robust representations for remote sensing (RS) understanding tasks. GeRSP combines supervised and self-supervised pre-training methods to simultaneously learn general and domain-specific knowledge. The framework includes two pre-training branches: a self-supervised branch for learning domain-related features from unlabeled RS images and a supervised branch for general knowledge learning from labeled natural images. GeRSP uses a teacher-student architecture to enhance the learning of both general and domain-specific knowledge, resulting in a powerful pre-trained model for deep learning initialization. The framework is evaluated on three downstream tasks: object detection, semantic segmentation, and scene classification. Experimental results show that GeRSP effectively improves the performance of remote sensing downstream tasks. The method addresses the domain gap between RS images and natural images by leveraging the rich knowledge of natural images, leading to better generalization and transferability. GeRSP outperforms existing pre-training methods in terms of performance and robustness, demonstrating its effectiveness in remote sensing tasks. The code and pretrained models are available at https://github.com/floatingstarZ/GeRSP.This paper proposes a novel remote sensing pre-training framework called Generic Knowledge Boosted Remote Sensing Pretraining (GeRSP) to learn robust representations for remote sensing (RS) understanding tasks. GeRSP combines supervised and self-supervised pre-training methods to simultaneously learn general and domain-specific knowledge. The framework includes two pre-training branches: a self-supervised branch for learning domain-related features from unlabeled RS images and a supervised branch for general knowledge learning from labeled natural images. GeRSP uses a teacher-student architecture to enhance the learning of both general and domain-specific knowledge, resulting in a powerful pre-trained model for deep learning initialization. The framework is evaluated on three downstream tasks: object detection, semantic segmentation, and scene classification. Experimental results show that GeRSP effectively improves the performance of remote sensing downstream tasks. The method addresses the domain gap between RS images and natural images by leveraging the rich knowledge of natural images, leading to better generalization and transferability. GeRSP outperforms existing pre-training methods in terms of performance and robustness, demonstrating its effectiveness in remote sensing tasks. The code and pretrained models are available at https://github.com/floatingstarZ/GeRSP.