Quantification of mineral end-members in lunar spectra through deep learning
The quantification of common lunar surface minerals through spectral unmixing is the primary objective of this thesis. Hyper spectral unmixing for sub-pixel identification of surface materials is one of the most challenging problem in the field of the planetary science. Spectral unmixing is the process of recovering the proportions of end-members that make up the complete spectra at each pixel of the image. There are different models available for spectral unmixing. The models can be data driven, such as neural networks, principal component analysis, fuzzy classifiers, etc.; or they can be physics based, such as radiative transfer, bilinear models, linear mixing models, etc. Recent studies shows that the hybrid models for linear and non-linear spectral unmixing of the hyper spectral data are increasingly using artificial neural networks. Neural networks are widely used in the area of remote sensing and hyper spectral unmixing problems due to their ability to recognize the complex patterns in higher dimension hyper spectral data. Artificial neural networks and Hapke's radiative transfer model along with Lunar Soil Characterization Consortium (LSCC) data has been used in this study for optimizing the model parameters and subsequently estimating the endmember abundances from lunar soil spectra. Hapke's adiative transfer model is one of the most widely used physics based model for characterizing regolith reflectance properties. This work is partitioned majorly in three parts: Optimization of the Hapke's radiative transfer model parameters; Generating the data set for the training of the neural network model, and Neural network model implementation. The Nelder mead optimization algorithm which is robust optimization technique for the multidimensional problems, has been used for the optimization of Hapke model parameters. This method finds the extreme values without finding derivative of the given function. The recurrent neural networks are used for the quantification of the minerals and it works well for predicting the abundances of up to 3 minerals .
- M Tech Dissertations