Enhancing Water Mask Extraction Through Image Recolorization and Sun Glint Mitigation

Duration of work: 4 months
Completion: March 2025
Supervisor: Dr.-Ing. Lida Asgharian Pournodrati
Prüfer: Prof. Dr.-Ing. Uwe Sörgel

Photogrammetry has emerged as a pivotal technology across numerous scientific and industrial domains, enabling accurate three-dimensional (3D) reconstruction from two-dimensional (2D) imagery. It is widely utilized in fields such as topographic mapping, environmental monitoring, agricultural planning, urban planning, infrastructure inspection, and hydrological studies due to its cost-effectiveness, precision, and the detailed spatial information it provides. Recent advancements in Unmanned Aerial Vehicle (UAV) technology have significantly expanded photogrammetric applications by offering high-resolution imagery, improved maneuverability, and flexible deployment capabilities. Despite these technological advancements, water surface photogrammetry remains a complex task, primarily due to image artifacts introduced by reflections and sun glint.

Reflection and sun glint on water surfaces occur due to environmental and weather conditions such as the time of day, sun angle, and intensity of sunlight at the time and location of aerial image capture. When sunlight directly reflects off the water surface, it causes bright spots and glare in aerial imagery. These phenomena severely degrade image quality, impairing the accuracy and reliability of water mask extraction processes critical for subsequent photogrammetric operations. Sun glint and reflections significantly hinder feature-matching techniques such as Structure-from-Motion (SFM) and Multi-View Stereo (MVS), consequently leading to erroneous or incomplete digital elevation models (DEMs). Such inaccuracies have substantial implications for hydrological modeling, environmental monitoring, and spatial analysis, compromising the validity and reliability of scientific assessments and decision-making.

Current methodologies addressing sun glint and water surface reflections predominantly include polarization filters, temporal minimum filtering, manual editing of point clouds, and multi-image fusion techniques. Each of these traditional methods has considerable limitations. Polarizing filters often provide inconsistent performance due to varying solar angles and changing weather conditions, while temporal filtering requires multiple passes, increasing the computational burden and limiting real-time applicability. Manual editing of point clouds remains costly and impractical for large-scale applications. Consequently, these limitations highlight a need for automated, efficient, and reliable techniques capable of mitigating sun glint and reflections directly from single aerial images, thereby facilitating robust water mask extraction.

Water surface photogrammetry is notoriously challenging due to image artifacts introduced by sun glint and reflections, which degrade key processes such as feature matching in Structure-from-Motion and Multi-View Stereo. This thesis proposes a deep learning–based framework to enhance water mask extraction by mitigating these artifacts directly within single UAV-captured images. Leveraging a Coarse-to-Fine U-Net pipeline in conjunction with Transformer blocks, the method progressively refines global color and local details across multiple resolution levels. 

Experimental evaluations were conducted on a combined dataset comprising both general reflective-surface images and a custom UAV-based water surface collection. Quantitative metrics Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) indicate that injecting Transformers into the Coarse-to-Fine architecture markedly improves reflection removal over baseline implementations. In addition to visual quality improvements, the method offers a more robust starting point for downstream photogrammetric workflows, particularly in generating accurate and complete 3D reconstructions of water-dominated scenes. 

While this study confirms the potential of combining Coarse-to-Fine U-Net designs and Transformers for sun glint mitigation, limitations remain regarding ground truth preparation, dataset diversity, and computational demands. Future work could include assembling larger, more varied UAV datasets, and integrating the proposed approach into end-to-end 3D reconstruction pipelines.