Friday 07 March 2025
Clouds are a nuisance for remote sensing researchers, obscuring valuable data and hindering their ability to study the Earth’s surface. In recent years, scientists have turned to fusion techniques to combine satellite imagery from different sensors and wavelengths in order to remove clouds from images. However, these methods often rely on complex algorithms and require extensive processing power.
A new approach has been proposed by researchers that uses a two-flow residual network (TFRN) to fuse polarimetric synthetic aperture radar (PolSAR) and optical data. This method leverages the strengths of both sensors to produce high-quality images with minimal cloud cover.
The TFRN is based on a deep learning architecture that uses cross-connection blocks to enable information sharing between PolSAR and optical images. The network is trained using a large dataset of paired images, where each image pair consists of a cloudy PolSAR image and its corresponding clear-sky optical counterpart.
During training, the TFRN learns to extract features from both types of images and combine them in a way that minimizes cloud cover while preserving surface details. The resulting images are then refined using an attention mechanism that selectively applies processing steps based on the importance of each pixel.
The researchers evaluated their approach using a dataset of airborne PolSAR and optical images, comparing it to several state-of-the-art methods for cloud removal. Their results show that the TFRN outperforms these methods in terms of both quantitative metrics (such as peak signal-to-noise ratio and structural similarity index) and visual quality.
One of the key advantages of this approach is its ability to handle complex terrain features, such as forests and urban areas, which are often challenging for cloud removal algorithms. The TFRN’s use of PolSAR data provides valuable information about the scattering properties of these surfaces, allowing it to better distinguish between clouds and surface features.
The authors suggest that their method could be particularly useful in applications where high-resolution imagery is required, such as monitoring crop health or tracking urban development. They also note that the approach is easily extensible to other types of sensors and data sources, making it a versatile tool for remote sensing researchers.
Overall, this new method demonstrates the potential for deep learning-based approaches to improve cloud removal in remote sensing applications. By leveraging the strengths of multiple sensor modalities, researchers can produce high-quality images with minimal cloud cover, enabling more accurate analysis and decision-making.
Cite this article: “Removing Clouds from Remote Sensing Images Using Deep Learning”, The Science Archive, 2025.
Cloud Removal, Remote Sensing, Deep Learning, Tfrn, Polsar, Optical Data, Image Fusion, Cloud Cover, Surface Details, Terrain Features.
