Deep learning has received intensive research focus in recent years. It has been applied to many fields including remote sensing classification. Convolution neural network is one of the important deep learning network architectures and has solid success in computer vision. In this study convolution neural network with different receptive field dimensions are used to classify hyperspectral imageries. The overall classification accuracies of CNN-1D, CNN-2D, CNN-3D and their variations are compared with those of traditional classifiers, such as RBF-SVM and MLP neural network used in this study. The result indicates all CNN models in the current study outperform RBF-SVM and MLP. Among the CNN models examined in this study the best CNN model is CNN-2D with its convolution process on a compressed PCA image. The improvement of overall classification accuracies of CNN-2D+PCA vs RBF-SVM models are 15.5% and 8.8% for Indian Pines and Pavia University hyperspectral imageries, respectively.
深度學習是人工智慧近幾年的研究熱點,並已經開始應用於遙測影像的地物辨識。卷積神經網路是深度學習的主要架構之一,其在電腦視覺應用取得十分傑出的成果;本研究將考慮不同神經網路感受域條件下之卷積神經網路,應用於高光譜影像的地物辨識,研究將分析一維、二維、三維卷積神經網路架構的整體辨識率,並與傳統支持向量機與類神經網路進行比較。研究結果顯示,在整體地物辨識率方面,所有架構之卷積神經網路均優於傳統支持向量機與類神經網路,這其中採用二維卷積神經網路加上PCA壓縮影像的效果最佳,其與支持向量機的結果比較,在Indian Pines高光譜圖資的地物辨識率增加15.5%,而在Pavia University高光譜圖資上地物辨識率增加8.8%。