[Colloquium] Miao He MS Presentation/Jun 1, 2023

[Megan Woodward]

This is an announcement of Miao He's MS Presentation
===============================================
Candidate: Miao He

Date: Thursday, June 01, 2023

Time:  1 pm CST

Remote Location: https://uchicago.zoom.us/j/96319028796?pwd=Mm9NUFptd0dIMGRyMXNmbUw1dXBxUT09

M.S. Paper Title: Deep Fair Partition for Hyperspectral Image Classification

Abstract: Over the last decade, hyperspectral images (HSIs) have attracted considerable attention in the remote sensing community, making HSI classification and segmentation a major concern. Convolutional neural networks (CNNs) have been widely used in HSI classification and segmentation due to their effectiveness and efficiency in feature extraction and expression. However, traditional patch-wise classification methods may easily lead to information leakage, rendering overoptimistic results and making it imperative to address data leakage issues to ensure classification performance. In this paper, we propose a novel data partitioning strategy along with a lightweight, attention-aided CNN-based model for HSI classification without any information leakage. Not only does the proposed data partition method avoid information leakage, but it also generates balanced data splits even if the training samples are limited and datasets are imbalanced. Though our model only exploits spectral features instead of extracting spectral-spatial information, it is very effective and efficient since there is abundant spectral information; moreover, the spatial information is not easy to utilize effectively due to the limited availability of training samples and low spatial resolution in datasets. Experimental results, collected by applying the proposed novel data partitioning scheme and attention-aided CNN networks to three standard hyperspectral datasets, outperforms the state-of-the-art CNN-based models without any information leakage.

Advisors: Rick Stevens

Committee Members: Rick Stevens, Ian Foster, and Fangfang Xia