Data-specific activation function learning for hyperspectral image classification

Hyperspectral image (HSI) classification plays an important role in lots of HSI analysis related tasks. In recent years, deep learning based methods draw much attention for HSI classification due to the powerful representation ability. The activation function is essential for deep learning based methods since it introduces nonlinearity into the network. However, existing activation functions such as Sigmoid and ReLU are pre-defined, which are handcrafted and general for any kinds of data. If they fit well to a specific dataset and thus lead to best classification result is seldom studied. In this paper, we propose to learn a specific data-driven function called data-specific (DS) activation function for HSI classification. Instead of using a hand-crafted function, we propose an HSI oriented activation function generation strategy, in which neural network (NN) architecture is utilized to learn the activation function suitable for HSI classification. Experiment results demonstrate the effectiveness of the learned activation function for HSI classification.