A Study of Deep Learning Approaches and Loss Functions for Abundance Fractions Estimation

Spectral unmixing is one of the post processing operation of hyperspectral image processing. In general, it is observed that the traditional algorithms are not efficient enough to estimate the abundance fractions of endmembers. In recent years, neural network based approaches are competent enough to perform the complex operations for remote sensing applications. After the origination of deep learning, number of deep learning approaches have been proposed and trained using different loss functions for performing several complex operations. Hence, it is necessary to select suitable deep learning model and loss function for the application(s). The objective of this work is to study the suitability of the model which affects the performance of the unmixing operation. In this work, estimation accuracy of different deep learning models and loss functions for spectral unmixing operation using hyperspectral data have been studied. In this work, five deep learning models are implemented and trained using four loss functions. Evaluation of proposed study has been carried using available real hyperspectral datasets. Hence, It is observed from the study that the different deep learning models and loss functions affect the estimation accuracy of spectral abundance fractions. In this work, parallel convolution 1-D model has been performed best among the implemented approaches for estimation of abundance fractions.