Abstract:

Hyperspectral images (HGs) are three-dimensional (3B) imaging cubes with two spatial and one spectral size. The development of deep learning methods has had a significant impact on HG classification. Especially the methods based on the evolutionary nervous network (ESA) are more interested in this field. In this study, we take advantage of the deep learning architectures of LeNet5, AlexNet, VGG16, GoogleNet and ResNet50, which are among ESA’s successful examples for HG classification problems. When using these architectures, we are using a 3B ESA-based hybrid approach. Because, using 3B ESA, spectral-spectrum properties are extracted simultaneously. In this case, the spectral-spectrum-based deep learning architecture achieved increases the classification accuracy of HGs. However, in the recommended model, basic component analysis (TBA) is used as a pre-processing technique for optimal band extraction from HGs. After the application of the TBA, 3B cubes are obtained with neighbourhood extraction and are given to the entrance of deep learning architectures. 3B ESA-based deep learning architectures have been used by Indian pines, Salinas, Botswana and HyRANK-Loukia data to compare the classification performance. As a result of the implementation, the best classification accuracy was obtained with the Indian pines data VGG16, ResNet in Botswana data50, VGG in HyRANK-Loukia data16, LeNet5 and VGG16 in Salinas data.

Keywords: