Inhaltszusammenfassung

In the realm of remote sensing, the quality variation of satellite images poses a substantial challenge across numerous application fields. Environmental monitoring, urban planning, agricultural management, and disaster response depend on high-quality data for accurate results. This paper presents a novel approach to address this issue by employing advanced super-resolution techniques. Real-ESRGAN, a state-of-the-art super-resolution convolutional neural network and known for its proficiency,...In the realm of remote sensing, the quality variation of satellite images poses a substantial challenge across numerous application fields. Environmental monitoring, urban planning, agricultural management, and disaster response depend on high-quality data for accurate results. This paper presents a novel approach to address this issue by employing advanced super-resolution techniques. Real-ESRGAN, a state-of-the-art super-resolution convolutional neural network and known for its proficiency, is trained on a large diversity of contents, ranging from people, handmade objects and environments to flora and fauna, and natural sceneries including underwater and dim light conditions. Leveraging the concept of transfer learning we acclimate this model to the inherent features of satellite images. We show that by shifting the knowledge from the pre-trained generalised Real-ESRGAN model towards satellite imagery, we obtain highly realistic and detail-rich results. This narrows the quality gap between low- and high-resolution imagery, thus facilitating more accurate analyses and comparisons. We train the model using Sentinel-2 and PlanetScope data to accommodate it to real-world satellite imaging conditions. Our methodology offers an accessible and cost-effective solution to obtain high-quality imagery from open data sources, approaching the standards of premium satellite solutions, without the need for highly specialised techniques.» weiterlesen» einklappen

Autoren

Klassifikation

DFG Fachgebiet:
Geophysik und Geodäsie

DDC Sachgruppe:
Technik