Classification of hyperspectral images by spectral-spatial dense-residual network

Hyperspectral images (HSIs) contain a significant amount of spectral and spatial information, together with underlying redundancy and noise, causing difficulty in HSI-processing tasks. State-of-the-art deep learning methods have obtained unprecedented performance in HSI classification and analysis. However, these architectures face challenges of declining accuracy and lengthy training times. We propose a framework to mitigate these issues, composed of a densely connected spectral block and preactivation bottleneck residual spatial block to separately learn spectral and spatial features. The spectral extraction block can involve more spectral features with the increase of the network depth, and it solves the problem of lengthy training time in traditional methods, and its densely connected structure achieves higher accuracy. In the spatial extraction block, we use the improved residual structure and introduce batch normalization and a parametric rectified linear unit before convolutional layers to preactivate the network, reducing parameters, and overfitting. In experiments using three classification approaches for comparison, it can be observed that even compared to the state-of-the-art method: spectral-spatial residual network for HSI classification, the proposed model shows improvements in accuracy of 0.49%, 0.19%, and 0.35% on the Indian Pines, University of Pavia, and Kennedy Space Center datasets, respectively. The experimental results reveal that the model obtains better classification results while effectively decreasing the training time. (C) 2020 Society of Photo-Optical Instrumentation Engineers (SPIE)