Frequency-Oriented Hierarchical Transformer for Advanced Satellite Image Dehazing and Downstream Task Preservation

Haze and smoke can significantly impair satellite images. Haze manifests in satellite images due to atmospheric scattering and poses difficulties in image analysis for remote sensing. The state-of-the-art dehazing methods do not consider key aspects such as frequency details, semantic edges, and subsequent tasks. In this paper, we propose a frequency-oriented Transformer connectionist framework with a task-consistency objective for satellite image dehazing. Our method comprises a frequency-aware attention mechanism, multi-scale feature reconstruction, and spatial and frequency fusion to achieve combined global and local scene restoration. Plus, we propose a task-consistency objective to restore the downstream usability of the dehazed images for remote sensing applications. Extensive experiments on paired datasets for remote sensing dehazing show that our method improves the fidelity and details of the restored images and achieves more reliable results on thin, moderate, and thick haze. Our proposed method aims to provide robust and useful preprocessing for Earth observation equipment.