BLIND NONLINEAR UNMIXING FOR INTIMATE MIXTURES USING HAPKE MODEL AND CNN

Abstract

We propose a blind nonlinear unmixing technique for intimate mixtures. We use the Hapke model and a fully convolutional neural networks (HapkeCNN). The proposed loss function contains 1) A quadratic term based on the Hapke model, 2) reconstruction error, and 3) a minimum volume term. The first term captures the nonlinearity, the second ensures the fidelity of the reconstructed reflectance, and the latter term exploits the geometrical information. The proposed method is evaluated using a simulated and a real datasets. We compare the results of endmember and abundance estimation with bilinear, multilinear, nonlinear, and projection-based linear unmixing techniques. The experimental results confirm that HapkeCNN considerably outperforms the state-of-the-art nonlinear approaches in terms of spectral angle distance and root mean square error. HapkeCNN was implemented in Python (3.9) using PyTorch as the platform for the deep network and is available online: https://github.com/BehnoodRasti/HapkeCNN.