Convolutional Neural Networks and Long Short-Term Memory for Hyperspectral Classification and Time Series Prediction

Photo Juan Rochac

PhD Candidate Juan Rochac

By: Juan F. Ramirez Rochac

PhD Advisor: Dr. Nian Zhang

Wednesday, December 14, 2022 at 2:00 PM

PhD Committee

Dr. Nian Zhang  PhD Advisor, UDC

Dr. Paul Cotae  Committee Chair, UDC

Dr. Lily Liang Member UDC

Dr. Xiang Chen External Member George Mason University

Dr. Laleh Najafizadeh External Member from Rutgers University

Defense abstract

PosterMachine Learning and Deep Learning (ML&DL) are areas of active research with multiple expert-level performing algorithms, yet additional research is needed to reach this level of accuracy on highly imbalanced, non-linear datasets. This is evident in the presence of environment data, such as remotely-sensed hyperspectral images and time series data that describe streamflow and surface-water quantity. State-of-the-art approaches are no longer reliable on these types of environment data and new classification and prediction models need to be explored due to the increasing imbalance and nonlinearity of hyperspectral classifiers and streamflow predictors. The benefit of accurate hyperspectral classification impacts environment monitoring, mineral exploration, and remote sensing while the benefit of accurate streamflow prediction impacts flood and drought management. Moreover, in places with legacy infrastructure, updated monitoring systems and unreliable forecasting frameworks, state-of-the-art, CNN-based and LSTM-based models suffer due to the presence of imbalanced and noisy data. In this study, we propose different approaches, as follows:

(i) The implementation of context-based feature augmentation (CFA) to tackle highly imbalanced data in hyperspectral classification using deep convolutional neuronets;

(ii) The implementation of Scharr-based adaptive filtering (SAF) to deal with non-linearities in time series prediction using deep recurrent neuronets; and finally,

(iii) The application of CFA and SAF to improve performance accuracy in hyperspectral classification and time series prediction using real-world environment datasets.

Hyperspectral datasets are collected from Telecommunications & Remote Sensing Laboratory (TLC&RS) and Time series datasets from the National Water Information System (NWIS). To implement the approaches, we use Python.


Webex Link

Hosted by Cotae, Paul

Wednesday, Dec 14, 2022 2:00 pm | 3 hours | (UTC-05:00) Eastern Time (US & Canada)

Meeting number: 2633 712 3995

Password: yNagy3pM7H3

Join by video system


You can also dial and enter your meeting number.

Join by phone


+1-415-655-0003 United States Toll

Access code: 263 371 23995

Join by phone

+1-415-655-0003 United States Toll

Access code: 263 071 92520

PhD in Computer Science and Engineering

Defense Presentations