運用生成對抗網路產製超解析衛星影像之先期評估

The Preliminary Evaluation of Generating Super-Resolution Satellite Images Using Generative Adversarial Networks

10.6574/JPRS.202409_29(3).0004

張庭韶(Ting-Shao Chang)；蔡富安(Fuan Tsai)

超解析 ； 衛星影像 ； 生成對抗網路 ； 深度學習 ； Super Resolution ； Satellite Imagery ； Generative Adversarial Network ； Deep Learning

航測及遙測學刊

29卷3期（2024 / 09 / 01）

177 - 189

繁體中文;英文

本研究探討使用生成對抗網路（GAN）模型提升衛星影像的空間解析度，以解決因拍攝角度、天氣狀況及感測器限制導致的解析度下降問題。研究使用中、高及超高解析度影像進行降解析處理，並透過GAN進行訓練生成超解析影像。GAN訓練過程中，生成器負責將低解析度影像重建為高解析度影像，判別器則區分生成影像與真實影像的差異。研究亦使用VGG-19預訓練模型進行特徵提取，提升生成影像品質。結果顯示，隨訓練次數增加，影像細節變得更為清晰及銳利化，且在結構相似性指標上優於傳統方法。然而，影像生成過程中出現色彩偏移及偽影現象。為改善此問題，建議進行更深層次訓練或使用後處理技術，並優化模型架構，如移除Batch Normalization。綜上所述，GAN模型具有提升衛星影像解析度的潛力，惟仍尚有影像色彩偏移及僞影問題，未來可針對模型穩定性和影像後處理進行優化。

This study explores the use of Generative Adversarial Networks (GAN) to enhance the spatial resolution of satellite images, addressing resolution degradation caused by factors such as off-nadir angles, weather conditions, and sensor limitations. The research utilizes medium, high, and very-high-resolution images, applying downsampling and training the GAN to generate super-resolution images. During GAN training, the generator reconstructs low-resolution images into super-resolution ones, while the discriminator distinguishes between generated and real images. The study also utilizes the VGG-19 pre-trained model for feature extraction to improve image quality. Experimental results show that image details become sharper as training progresses, and the GAN outperforms traditional methods in terms of structural similarity. However, issues like color shifts and artifacts emerged during image generation. To address these problems, the study recommends deeper training, post-processing techniques, and model optimizations such as removing Batch Normalization. Overall, while GAN models show potential for enhancing satellite image resolution, further improvements are needed to resolve color shifts and artifacts, focusing on model stability and post-processing.