04/27/2021
By Dalila Megherbi

Vusal Pasha will defend his MS thesis in Computer Engineering, titled " Hyperspectral Image Classification/Prediction via Deep Learning with Additive Noise," on Tuesday, May 18, 2021, at 10 a.m.

Location: This will be a virtual defense via Zoom. Those interested in attending should contact Committee Chair Dalila_Megherbi@uml.edu and vusal_pasha@student.uml.edu at least 24 hours before the defense to request access to the meeting.

Committee Chair: D. B. Megherbi, ECE Department, (CMINDS), UML, (Thesis Advisor)

Committee Members:

  • Kanti Prasad, ECE Department, UML
  • Xuejun Lu, ECE Department, UML

Abstract:

Hyperspectral imaging is a spectral sensing technique to obtain data cubes consisting of images of a given scene recorded at multiple spectral (250 plus) bands within the electromagnetic spectrum. Recent studies have focused on utilizing hyperspectral data to perform classification/prediction with various Deep Learning (DL) based-hyperspectral image (HSI) classification methodologies in various applications. This includes applications in military defense, archeology, medical diagnosis, analysis of crime scene detail, forensics, quality control, and mineralogical mapping of surfaces, to name a few. Due to the high dimensionality of the data, lack of effective dimensionality reduction (DR) techniques, and noise across non-visible spectral bands, many of such HSI classification/prediction systems were limited in performance.

This thesis focuses on enhancing HSI classification/prediction while coping with the existing limitations by introducing a new DL-based methodology, which outperforms some state-of-the-art methods and improves the 3D-DenseNet architecture. It also analyses additive noise on the proposed predictive model and the external validating test data. Specifically, we (a) propose a new 3D-HSI-DenseNet architecture that achieves relatively better classification/prediction accuracy than the previously developed 3D-DenseNet model while reducing the total number of parameters by nearly 1.5 times; (b) demonstrate that the proposed model has a faster training and testing computational times; (c) show that the proposed predictive model is stable and has a smaller variation in prediction while achieving higher prediction accuracy than related state-of-the-art methodologies; (d) experimentally show that Network-In-Network (NIN) can be used for DR of HSI and it is as effective as the PCA for DR; (e) show the change in HSI classification/prediction performance of the proposed model trained with and without Gaussian noise while introducing different levels of noise into the test data. We show the potential value of the proposed methodology in comparison to four selected DL state-of-the-art performing models from the literature. We used Keras as a framework to build, train, and test our model on a V100-SXM2 GPU.

All interested students and faculty members are invited to attend the online defense via remote access.