Comparison of Small-Sized Deep Neural Network Models for Hyperspectral Image Classification

Abstract

Many different methods related to deep learning are used in a broad scope in areas such as identification, clustering, and classification of 1D, 2D, and 3D data. In contrast to traditional images, deep learning studies are also performed on hyperspectral images, which are structurally different by the hundreds of bands it contains. Although hyperspectral images are data in 3D format, they can be analyzed separately in 1D, 2D and 3D formats, which is advantageous for deep learning research. Although the presence of so many different types of information provides diversity for studies, on the other hand, it significantly affects the values and working times of the results. In the scope of this study, the classification of hyperspectral images using convolutional neural networks, a subtitle of deep learning, was carried out. The classification results obtained by convolutional neural networks with different characteristics and changing parameters were examined. In this way, the effect of other methods on the classification results was also discussed. Within the scope of the whole study, all methods were tested on Indian Pines, Salinas and Pavia University datasets, frequently preferred in the literature. The results obtained with the methods we used were compared with the results of other studies in the literature. It has been confirmed that the network models with the 3D convolution layer mentioned in the literature and our studies have better performance results. Furthermore, although our proposed models contain fewer parameters, unlike the models in the literature, they almost approach or exceed the other models’ results.