Patch-Wise Hyperspectral Image Classification Using Composite 3d-2d Convolutional Neural Network Feature Hierarchy

Abstract

Recent advances in hyperspectral imaging have increased the use of convolutional neural networks for classification. However, using only a 2D or 3D convolutional neural network necessarily involves high computational complexity and does not effectively exploit spectral spatial features. Traditional classifiers, such as Support vector machine (SVM) provide adequate classification accuracy; however, SVM cannot properly classify data when classes overlap. A large number of training samples are also required. To address the shortcomings of such traditional classifiers, a composite 3D-2D convolutional neural network feature hierarchy is developed. The proposed method exploits both the spectral and spatial features of the images very well, while the 2D convolutional layer aids in the spatial feature extraction. The dataset is divided into 3D cubes or patches, and the performance of the patches is compared using sigmoid and SoftMax loss functions, along with adaptive motion estimation (ADAM) and stochastic gradient decent (SGD) optimization. When compared to traditional classifiers such as SVM, the proposed method achieves high classification accuracy. The effect of patch sizes and training sample sizes on classification accuracy is studied. The purpose of this study is to help and encourage other researchers to conduct more research in this topic.