康立威許志仲Kang, Li-WeiHsu, Chih-Chung何祐豪Ho, Yu-Hao2025-12-092025-08-062025https://etds.lib.ntnu.edu.tw/thesis/detail/47db792f4697afb9fc96f6eef450574c/http://rportal.lib.ntnu.edu.tw/handle/20.500.12235/125062在衛星遙測領域 (Satellite Remote Sensing) 中高解析高光譜影像 (HRHSI, High Resolution Hyperspectral Image) 的獲取與生成一直是熱門的研究議題之一，原因是使用高解析高光譜影像能更加提升遙測物件偵測與辨識任務的準確率與可靠性。而使用低解析度高光譜影像 (LRHSI, Low Resolution Hyperspectral Image) 以及高解析度多光譜影像 (HRMSI, High Resolution Multispectral Image) 進行融合是常見且優秀的方法，此方法可以同時結合高光譜影像豐富的光譜資訊與多光譜影像的空間資訊以融合生成高品質的高解析高光譜影像。本論文中我們提出一個基於Transformer高光譜與多光譜影像融合任務用的類神經網路模型，此模型稱為HSSDCT，其透過Residual Dense Connection能降低訓練時資訊的損失並且有效地再利用特徵圖 (Feature Map)，並使用SSC (Spatial-Spectral Correlation) 的架構以提升模型空間與光譜特徵的提取能力。此外，此模型採用階層式架構建立模型，透過由小到大改變每層Transformer的窗口大小 (Window Size) 以達到提取局部資訊與維持長程依賴性 (Long-Range Dependencies) 的效果。實驗結果顯示，此方法能以適當的模型大小與參數量融合生成高品質的高解析高光譜影像。In the field of satellite remote sensing, the acquisition and generation of high-resolution hyperspectral images (HRHSI) has been one of the popular research topics. The reason is that the use of high-resolution hyperspectral images can further improve the accuracy and reliability of remote sensing object detection and identification tasks. The fusion of low-resolution hyperspectral images (LRHSI) and high-resolution multispectral images (HRMSI) is a reliable solution. This method can combine the rich spectral information of hyperspectral images and the spatial information of multispectral images to generate high-quality high-resolution hyperspectral images. In this paper, we propose a neural network model based on Transformer for hyperspectral and multispectral image fusion tasks. This model is called HSSDCT. It can reduce the information loss during training and effectively reuse feature maps through Residual Dense Connection, and use the SSC (Spatial-Spectral Correlation) architecture to improve the model's ability to extract spatial and spectral features. In addition, this model uses a hierarchical architecture to build the model, by changing the window size of each layer of Transformer from small to large to achieve the effect of extracting local information and maintaining long-range dependencies. Experimental results show that this method can generate high-quality high-resolution hyperspectral images with appropriate model size and parameters.高光譜影像高解析度高光譜影像低解析度高光譜影像高解析度多光譜影像影像融合長程依賴性Hyperspectral imageHRHSILRHSIHRMSIimage fusionlong-range dependencies學術論文