Towards UAV Image Dehazing: A UAV Atmospheric Scattering Model, Benchmark, and Geometry-Aware Deep Unfolding Network

A research team has proposed a new physics-driven framework to remove haze from unmanned aerial vehicle imagery, addressing limitations in how existing models handle the non-uniform haze typical of high-altitude and pitched camera views [1]. The work, submitted on 15 Jun 2026, introduces the UAV Atmospheric Scattering Model (UASM) and a companion processing network called the Geometry-Aware Deep Unfolding Network (GP-DUN) [1]. The standard atmospheric scattering model, long used in image dehazing, assumes a uniform haze distribution that does not hold for UAV platforms, where flight altitude, viewing pitch, and vertical atmospheric stratification jointly influence how haze appears in an image [2]. UASM explicitly incorporates these geometric factors to characterize the spatially varying haze [1]. Image dehazing has been a critical research area for remote sensing and UAV applications because haze obscures surface detail and weakens structural information [3]. A 2024 review of the field noted that prior-based dehazing methods often perform well in one scenario but fail in others, while deep-learning approaches offer greater flexibility but still struggle with the unique challenges of UAV imagery [3]. The authors of the new paper argue that a core obstacle is the unavailability of real-world paired hazy and clean UAV images for training, combined with the inadequacy of classical models for UAV-specific haze patterns [1]. To overcome this, the team built GP-DUN around three modules. A Latent Geometry Estimator infers transmittance consistent with UAV imaging geometry. A Geometry-aware Gradient Descent Module embeds UASM into a data-fidelity term for physics-consistent updates. A Pooling-Expert Proximal Mapping Module learns an implicit prior to restore textures beyond what explicit physical modeling can achieve [1]. The researchers also constructed UASM-HazeSet, a benchmark that provides controllable paired synthetic data and 2,285 real UAV haze images for testing [1]. The demand for robust UAV dehazing is growing as drones are deployed for aerial photography, precision agriculture, infrastructure inspection, and environmental monitoring [9]. Other recent efforts have produced realistic haze datasets with in-situ smoke measurements aligned to aerial and ground imagery, such as the A2I2-Haze dataset, to evaluate both dehazing and object-detection algorithms [5]. A 2025 survey of remote sensing image dehazing categorized the field into a progression from handcrafted physical priors to data-driven deep restoration and hybrid physical-intelligent generation, finding that models with explicit transmission or airlight constraints can reduce color bias by up to 27 percent [4]. The authors report that GP-DUN consistently outperforms existing methods on both UASM-HazeSet and real UAV haze benchmarks [1]. They express hope that the benchmark will facilitate future research on UAV-oriented image restoration under adverse weather conditions [2].