Deep Learning Models for Small Sample Hyperspectral Image Classification

A detailed background view of Hyperspectral remote sensing shows that hyperspectral image classification (HSIC) is very useful in many real-life applications. The existing deep learning models' capacity to learn new task-specific knowledge for generalization from a limited data sample that encourages less training is lacking in hyperspectral image classification. Development in deep learning has shown that dense models have better classification performance. However, this rule is not suitable for small data sample problems, especially in an unsupervised learning HSIC. The computational cost required by the state-of-the-art complex deep learning models renders them infeasible in constraint resource application environments. Consequently, this paper presents an analytical review of deep learning models for small sample HSIC. The focus is placed on supervised, semi-supervised, and unsupervised learning. Meta-learning, Meta-reinforcement Learning, and Capsule Networks are realized to be promising for Small Sample Hyperspectral Image Classification.