Proceedings Article | 4 December 2024
KEYWORDS: Image processing, Satellites, Satellite imaging, Neural networks, Remote sensing, Signal to noise ratio, Education and training, Data processing, Earth observing sensors, Image quality
Deep learning networks have been widely used in remote sensing image processing, playing a significant role in ground-based intelligent processing, on-board intelligent processing, and terminal-based intelligent processing. However, traditional intelligent processing methods for ground, on-board, and terminal applications exhibit substantial differences in methodology, network architecture, and outcomes. This leads to issues such as large data transmission volumes and poor image quality when terminal users directly utilize on-board processed images. Designing neural networks with identical architectures and employing the same methods for collaborative intelligent processing on the ground, on-board, and at the terminal can effectively enhance the benefits of terminal users applying on-board processed images. This paper proposes a satellite-terminal collaborative processing method for remote sensing images based on neural networks, where both the satellite and ground terminals employ the same network architecture. The method involves learning from onboard processed image sample data on the ground, synchronizing the learned network parameter data to the satellite and terminal, and utilizing the same neural network for on-board and terminal intelligent processing, thereby achieving satellite-terminal collaborative intelligent processing for remote sensing images. By using this method to learn from onboard processed image sample data and testing a portion of the sample data with the network, the paper analyzes the test sample data. Experimental results indicate that this method can achieve data compression, reduce data volume, and alleviate satellite-terminal data transmission pressures after intelligent processing of on-board images. Furthermore, after terminal-based intelligent processing, the image denoising effect is significant, and the data quality is substantially improved.
