HIMOSA: Efficient Remote Sensing Image Super-Resolution with Hierarchical Mixture of Sparse Attention

Yi Liu, Yi Wan, Xinyi Liu, Qiong Wu, Panwang Xia, Xuejun Huang, Yongjun ZhangarXiv 2025 · DOI · arXiv

Introduction

High-resolution (HR) remote sensing imagery is critical for applications ranging from disaster response to urban planning. However, hardware limitations and bandwidth constraints often result in low-resolution (LR) captures. Single-image super-resolution (SISR) provides a computational path to reconstruct HR details from LR inputs. While Transformer-based models have set new benchmarks in SISR, their quadratic computational complexity ($O(N^2)$) often hinders their deployment in time-sensitive remote sensing tasks. HIMOSA (Hierarchical Mixture of Sparse Attention) addresses this by introducing a content-aware, sparse attention mechanism that maintains high performance with significantly reduced complexity.

Methods

The HIMOSA framework is designed to exploit the inherent redundancy and multi-scale patterns found in aerial and satellite imagery. The architecture consists of several key components that work together to balance efficiency and reconstruction quality.

The overall HIMOSA architecture featuring hierarchical blocks and sparse attention mechanisms.

1. Content-Aware Routing Sparse Attention (CARSA)

Inspired by the "Mixture of Experts" (MoE) concept used in large language models, CARSA avoids the $O(N^2)$ complexity trap by selecting only the top-$k$ most informative tokens for processing. Unlike traditional methods where tokens choose experts, HIMOSA employs an "Expert Choice" routing where experts select the most relevant tokens. This ensures a balanced workload across attention heads and focuses computational power on critical image features like edges and complex textures, shifting complexity to a leaner $O(k^2 + N)$.

Detailed view of the core modules: (a) Channel Attention, (b) CARSA with token selection, and (c) the ConvGLU block.

2. Hierarchical Window Expansion and Adaptive Sparsity

Remote sensing scenes contain repetitive structures at various scales, such as rows of houses or airport layouts. HIMOSA utilizes a series of layers where the attention window progressively expands. This allows the model to capture local textures in early stages and broader contextual relationships in later stages. To maintain efficiency as windows grow, the model automatically increases sparsity, ensuring that only the most salient information is processed even at larger scales.

3. Supporting Modules

The architecture is further enhanced by a Channel Attention module for global feature integration and a Convolutional Gated Linear Unit (ConvGLU). ConvGLU introduces spatial adaptability and non-linearity, which are essential for handling the diverse geometric structures present in remote sensing data.

Datasets and Evaluation

The model was rigorously evaluated on four prominent remote sensing benchmarks:

  • AID: A large-scale dataset used for training and detailed category-wise testing.
  • DOTA v2.0 & DIOR: Massive benchmarks focused on object detection in aerial imagery.
  • NWPU-RESISC45: A diverse dataset containing 45 distinct scene classes.

Key Results and Insights

HIMOSA achieves state-of-the-art (SOTA) results across all tested datasets. On the AID benchmark, it demonstrated a PSNR improvement of 0.06 dB and an SSIM boost of 0.0021 over previous leading models like SwinIR and HAT.

A critical advantage of HIMOSA is its efficiency. The HIMOSA-light variant achieved the fastest inference speed among lightweight models, processing a 256x256 patch in just 63.82 ms—a 27% improvement over the next best competitor. Qualitative analysis confirms that HIMOSA excels at reconstructing fine, repetitive textures, such as running track lines or parking lot grids, where other models often produce blurry or aliased artifacts.

Conclusion

HIMOSA represents a significant step forward in making high-quality super-resolution practical for remote sensing. By leveraging content-aware sparsity and a hierarchical attention structure, it breaks the traditional trade-off between computational speed and reconstruction accuracy. Future work will explore the application of these sparse attention mechanisms to other restoration tasks, such as de-noising and de-hazing in satellite imagery.

Summary updated

BibTeX
@misc{liu2025himosa,
  author = {Yi Liu and Yi Wan and Xinyi Liu and Qiong Wu and Panwang Xia and Xuejun Huang and Yongjun Zhang},
  title = {{HIMOSA: Efficient Remote Sensing Image Super-Resolution with Hierarchical Mixture of Sparse Attention}},
  year = {2025},
  eprint = {2512.00275},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.00275},
}
Figure