Behnood Rasti, Bikram Koirala, Paul Scheunders, and Pedram Ghamisi
Hyperspectral sensors capture images at high spectral resolution, which allows detecting unique color signatures of individual objects. Due to limited spatial resolution of the images and scattering of light, a pixel on a hyperspectral image generally contains a mixture of the spectra from several objects. To decompose the measured spectrum into pure spectra of its constituent materials, or endmembers, and their corresponding fractions, or abundances, unmixing of the image is required. Linear unmixing models assume that light rays interact with only one material before reaching the sensor. Linear unmixing techniques are used in remote sensing applications and are the focus of this work.
We introduce a new method called hyperspectral unmixing using deep image prior (UnDIP). The main goal is to boost the abundance estimation and to make the unmixing problem robust to noise. The method comprises a conventional geometric approach for endmember extraction and a new unmixing deep image prior for abundance estimation. Geometric endmember estimation initializes deep unmixing; the endmembers are included in the loss function and remain fixed. Deep image prior employs an unsupervised convolutional neural network (CNN) that does not need spectral signatures for training. It uses Gaussian noise as input and generates fractional abundances by iteratively minimizing an implicitly regularized loss function. As a result, CNN can be applied globally to the entire spatial domain of an image, which produces sharper abundance maps and enhances robustness to noise. The new method is evaluated on the simulated and real remote sensing data and shows considerable improvements compared to state-of-the-art methods.
The proposed method is implemented in Python (3.8) using PyTorch and is available on GitHub.
IEEE Transactions on Geoscience and Remote Sensing, 2021-03-31.