Lightweight Remote Sensing Image Dehazing via Dual-Stage Feature Calibration and Hierarchical Context Aggregation

Remote sensing images (RSIs) are frequently degraded by atmospheric haze, which introduces color distortion and contrast reduction, thereby impeding downstream applications. Existing models often struggle with non-uniform haze distributions, high computational costs, and the loss of local texture details. To address these challenges, this paper proposes a lightweight Feature Self-Recalibration Network (FSRNet) for efficient remote sensing image dehazing. FSRNet adopts a symmetric encoder–decoder architecture as its backbone and utilizes parameter-free pixel shuffle and unshuffle operations for multiscale feature resampling to preserve complex spatial details. The core of FSRNet lies in the specially designed Feature Self-Recalibration Module (FSRM), which consists of two key components: the Dual-Stage Feature Calibration Block (DFCB) and the Hierarchical Context Aggregation Block (HCAB). Specifically, the DFCB statistically splits features into informative and redundant parts, independently recalibrating them through a simplified channel attention mechanism to enhance representation in heterogeneous haze regions. Simultaneously, the HCAB integrates a non-local haze perception branch and a local detail enhancement branch in parallel, enabling the model to perceive global haze density while preserving fine-grained textures. Experimental results on multiple authoritative synthetic and real-world remote sensing datasets demonstrate that FSRNet achieves state-of-the-art dehazing performance. With only 0.865 M parameters and 8.622 G MACs, FSRNet strikes a superior balance between restoration quality and computational efficiency, making it highly suitable for real-time deployment on resource-constrained platforms.