Spectral Unmixing in Generative Space: 3D-GAN Based Approach

Spectral unmixing in planetary data encounters challenges like noise, limited samples, and high dimensionality, affecting its efficacy. Deep learning methods, while highly performant, often lack interpretability and generalizability. In this study, we comprehensively evaluate spectral unmixing techniques, considering accuracy, efficiency, interpretability, generalizability, noise robustness, and sample requirements. Our primary focus is on Generative Adversarial Networks (GANs), specifically comparing the nonlinear Polynomial Post-Nonlinear Mixing Model (PPNM) with our proposed 1D-GAN and 3D-GAN. In our GAN approach, one CNN discriminates inputs, and another generates fake inputs. They train adversarially, enhancing discriminative CNN's generalization, crucial with limited samples. We propose a 1D-GAN for spectral classification and a robust 3D-GAN for spectral-spatial tasks. Using generated adversarial samples for fine-tuning, we improve endmember abundance accuracy. We apply this method to hyperspectral image (HSI) obtained from the Imaging Infrared Spectrometer (IIRS) sensor onboard the Indian Space Research Organisation's (ISRO) Chandrayaan-2 mission yielding competitive results. Our GAN approach opens new opportunities in remote sensing for challenging spectral unmixing, revealing GAN potential in analyzing complex, nonlinear data.