Multisource Domain Transfer Learning Based on Spectral Projections for Hyperspectral Image Classification

Abstract

Hyperspectral image classification is an important topic for hyperspectral remote sensing with various applications. Hyperspectral image classification accuracy has been greatly improved with the introduction of deep neural networks, while the idea of transfer learning provides an opportunity to solve the problem even with the lack of training samples. In this paper, we propose an effective transfer learning approach for hyperspectral images, projecting hyperspectral images with different sensors and different number of bands into a general spectral space, preserving the relative positions of each band for spectral alignment, and designing a hierarchical depth neural network for shallow feature transfer and deep feature classification. The experiments show that the proposed method can effectively preserve the source domain features, especially for the scenarios with very few samples in the target domain, which can significantly improve the classification accuracy and reduce the risk of model overfitting. Meanwhile, this strategy greatly reduces the requirement of source domain data, using multi-sensor data to jointly train a more robust general feature model. The proposed method can achieve high accuracies even with few training samples compared to currently many state-of-the-art classification methods.