Graph-based Hyperspectral Unmixing Using Deep Convolutional Neural Networks

Abstract

Hyperspectral image unmixing (HU) is a remote sensing technique that has numerous applications in ecology, earth observation, and agronomy. Recently the research surrounding hyperspectral unmixing are using deep convolutional neural networks (CNNs). This helps to extract more detailed features of the image scene through multi-layer non-linear transformation and it has better performance than previous unmixing approaches. However, there are challenges that affect the performance of CNNs in the unmixing process. First, training the CNN with pixels which have different non-linearity mixing functions lead to miss-estimation of network parameters. Second, hundreds of spectral bands in the hyperspectral image (HSI) may cause the curse of dimensionality in data processing by the CNNs. Third, the sparsity of annotations is another concern in CNNs due to loss of detailed information. Fourth, over-fitting is another problem in Deep CNNs (DCNN) when the training data has a small number of endmembers. In this thesis, we will address some of the challenges of deep CNN applied in the unmixing process by proposing novel pixel-wise spectral unmixing methods. First, we use a graph model for the HSI to classify the pixels into homogeneous and heterogeneous classes based on minimizing the Structural Intervention Distance. We then train a dedicated DCNN for each class of pixels. In the next step, we propose another HU method called HU-DCNN-MF. Taking interference, distortion and noise into account, we design a bank of Matched Filters (MF) in the HU-DCNN-MF approach as a linear preprocessing stage before passing the data pixels through deep CNNs. The bank of matched filters improve the signal to interference, distortion and noise ratio (SIDNR) of the input data for the DCNNs. To consider the spatial correlations in HU, we develop a bank of Optimal Filters (OF) by using a graph model for the HSI. In this method called HU-DCNN-OF, we employ a bank of optimal filters cascaded with the DCNNs. We maximize the SIDNR of the HSI's whole spectral bands by the bank of optimal filters before processing them using DCNNs. Our experimental results show that the main 3 proposed methods outperform the conventional state-of-the-art pixel-wise deep CNN methods.