| 摘要: | 在地球環境觀測的領域，遙測影像是一個能提供大尺度之時間與空間觀測資料的利器。其中，高光譜影像的波段數為數十個至數百個波段，能夠提供目標地物的細微光譜特徵，並以此做為地物分類的依據。然而，在同樣能量強度的入射光源下，每個波段能分配到的能量比較少，使得高光譜影像的空間解析度被犧牲。超解析影像處理是一種電腦視覺領域的影像處理技術，目的在於提升影像的空間解析度，目前已被廣泛應用在各式各樣的領域，例如文字辨識、安全監控、手機的修圖軟體等等。近年來超解析影像模型的運算能力越來越強大，將這樣的技術應用於提升高光譜遙測影像的空間解析度，或許可以協助改善地物分類的準確度。
本論文聚焦在以卷積神經網路為架構的超解析影像模型，例如MS-LapSRN與EDSR，並比較這些超解析模型對於提升高光譜遙測影像之空間解析度的效果。本文使用Sentinel-2與SPOT-6/7光學衛星影像作為訓練及測試超解析模型的資料集，結果顯示MS-LapSRN_D5R8的模型在峰值訊噪比 (PSNR) 與結構相似性 (SSIM) 指標上有較好的表現。接著，再以高光譜遙測影像分類基準資料集，例如Indian Pines 資料集與Pavia University資料集，以最小歐幾里得距離法以及傳統類神經網路進行監督式純像元分類，並比較提升空間解析度之前與之後，在地物分類任務的表現。本研究針對兩組高光譜影像之實驗成果顯示，影像超解析不一定能提升高光譜影像分類的準確率，針對其影響的規律需要更進一步的研究來探討。;In the field of earth environment observation, remote sensing imagery is a powerful tool that can provide large-scale temporal and spatial observation data. Among them, hyperspectral images have hundreds of bands, which can provide subtle spectral characteristics of target objects and serve as the basis for land cover classification. However, under the same energy intensity of the incident light source, the amount of energy for each band is reduced, resulting in the sacrifice of spatial resolution of hyperspectral images. Super-resolution (SR) image processing is an image processing technology in the field of computer vision that aims to improve the spatial resolution of images. It has been widely used in a variety of fields, such as text recognition, security monitoring, and mobile phone photo editing software. In recent years, the computing power of super-resolution image models has been significantly improved. Applying such technology to improve the spatial resolution of hyperspectral remote sensing images might help improve the accuracy of land cover classification.
This study focuses on image super-resolution models that are based on convolutional neural networks, such as MS-LapSRN and EDSR, and compares the effects of these super-resolution models on improving the spatial resolution of hyperspectral remote sensing images. The optical satellite imagery, Sentinel-2 and SPOT-6/7, were used as training and testing dataset for SR model. The results show that MS-LapSRN_D5R8 model has better performance in peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). Furthermore, hyperspectral image classification benchmark datasets, Indian Pines dataset and Pavia University dataset, were used to compare the performance in land cover classification tasks before and after improving the spatial resolution. Minimum Euclidean distance (MED) and traditional neural network (NN) was used as supervised pure-pixel classification method. It was observed in the context of this study that image super-resolution does not necessarily improve the accuracy of hyperspectral classification, and further studies are needed to investigate this issue. |
