<p>Satellite imagery plays a crucial role in various applications, such as environmental monitoring, urban planning, and disaster response. Previously, Convolutional neural networks have played a crucial role in improving remote sensing scene classification by effectively capturing spatial features. However, they often struggle with long-range dependencies and complex spectral patterns in satellite images. Recently, transformer-based architectures, initially designed for natural language processing, have been adopted in computer vision to leverage self-attention mechanisms for capturing long-range dependencies. To effectively extract the long-range dependencies and complex patterns, this paper proposes a novel method combining dense and sparse attention mechanisms to enhance remote sensing scene classification in different scales. The dense attention mechanisms capture comprehensive connections within satellite images, followed by sparse attention mechanisms applied at both tile and patch scale levels. This multi-scale attention strategy allows the model to efficiently focus on important spatial and spectral features while managing computational complexity. Extensive experimental evaluations on three benchmark datasets, such as Merced, AID, and Optimal-31, demonstrate significant improvements in classification accuracy, achieving 98.9, 97.5, and 97.8, respectively. Experimental results show that the proposed framework outperforms CNN and Transformer baselines, enhancing the accuracy and robustness of remote sensing scene classification for geospatial analysis.</p>