Three-dimensional densely connected convolutional network for hyperspectral remote sensing image classification

Abstract. Hyperspectral remote sensing images (HSIs) are rich in spatial and spectral information, thus they help to enhance the ability to distinguish geographic objects. In recent years, great progress have been made in image classification using deep learning (such as 2D-CNN and 3D-CNN). Compared with traditional machine learning methods, deep learning methods can automatically extract the abstract features from low to high levels and convert the images into more easily recognizable features. Most HSI classification tasks focus on spectral information but often ignore the rich spatial structures in HSIs, leading to a low classification accuracy. Moreover, most supervised learning methods use shallow structures in HSI classifications and hence exhibit weak performance in finding sparse geographic objects. We proposed to use the three-dimensional (3-D) structure to extract spectral–spatial information to build a deep neural network for HSI classifications. Based on DenseNet, the 3D densely connected convolutional network was improved to learn spectral-spatial features of HSIs. The densely connected structure can enhance feature transmission, support feature reuse, improve information flow in the network, and make deeper networks easier to train. The 3D-DenseNet has a deeper structure than 3D-CNN, thus it can learn more robust spectral–spatial features from HSIs. In fact, the deeper network structure has a regularized effect, which can effectively reduce overfitting on small sample datasets. The network uses HSIs instead of feature engineering as input data and is trained in an end-to-end manner. The experimental results of this model on the Indian Pines datasets and the Pavia University datasets show that deeper neural networks further improve the classification of complex objects, especially in the areas where geographic objects are sparse. It effectively improves the classification accuracy of HSIs.