Improving stacked-autoencoders with 1D convolutional-nets for hyperspectral image land-cover classification

Deep learning opened new possibilities for remote sensing image analysis using multiple neural nets layers. We introduce a hybrid pixel-based model that allows improving the unsupervised training with stacked autoencoders (SAE) by inserting convolutional neural networks (CNN) in the encoding and decoding steps. Inclusion of the convolution in the encoding and decoding steps allows a feature-based description of the pixel’s hyperspectral signature, suitable to perform an initial unsupervised classification. As one-dimensional (1D) filters are applied, the processing effort is lower than when using two-dimensional-CNN. Finally, to adapt the classifier to the desired classes, the parameters of the net are adjusted using training samples and fine-tuning followed by logistic regression using the softmax activation function. This combination explores the potential of both, autoencoders (AE) and convolutional nets, providing an alternative for the classification of hyperspectral data. To evaluate the performance of the proposed approach, it was compared to traditional machine learning algorithms such as support vector machine, artificial neural networks, CNN, and SAE. The results show that the use of the SAE-1DCNN method is more effective in terms of hyperspectral classification accuracy and more efficient in computational complexity and that it can be an alternative for hyperspectral data classification.