3D Convolutional Neural Network for Hyperspectral Image Classification Using Generative Adversarial Network

Abstract

Recently, hyperspectral image (HSI) classification has become a hot spot in the field of machine learning. Unlike traditional black and white dual-channel images or R, G, B three-channel images, HSI has multiple channels in the spectral dimension, one channel captures light of a specified wavelength. Hyperspectral data sets are usually three-dimensional. In order to extract the spectral spatial features in such images and differentiate images from unknown classes, We have introduced domain adaptation technology, which is a well established technique for appling an algorithm trained in one or more "source domains" to a different (but related) "target domain", and proposed a framework based on Three-dimensional Convolutional Neural Network (3D-CNN), combined with Generative Adversarial Networks (GAN) to deal with unknown classes hyperspectral image classification. Our framework can achieve high-precision classification on hyperspectral images. Compared with other deep learning based methods, Our model has advantages in many aspects, such as universality for different hyperspectral data sets, and requiring fewer parameters and epochs during training. We use datasets of Pavia University scene to evaluate the adaptability of our method on the open set domain. The classification accuracy of the proposed method for unknown categories is 87.16% on PaviaU and 99.43% on Salinas.