Publication
Preprocessing Ground-Based Hyperspectral Image Data for Improving CNN-based Classification
Andreas Schliebitz; Heiko Tapken; Martin Atzmueller
In: Michael Leyer; Johannes Wichmann (Hrsg.). Lernen, Wissen, Daten, Analysen (LWDA) Conference Proceedings, Marburg, Germany, October 9-11, 2023. GI-Workshop-Tage "Lernen, Wissen, Daten, Analysen" (LWDA-2023), October 9-11, Marburg, Germany, Pages 399-413, CEUR Workshop Proceedings, Vol. 3630, CEUR-WS.org, 2023.
Abstract
Complex data–like hyperspectral image data–requires adequate preprocessing methods for tackling the issue of data quality, as a prerequisite for further machine learning approaches like deep learning. This paper addresses preprocessing in the context of ground-based hyperspectral image data: It presents novel preprocessing methods, and proposes a comprehensive preprocessing pipeline for handling complex hyperspectral image samples. Multiple preprocessing pipelines are applied on a set of hyperspectral images in the context of image classification, analyzing which preprocessing algorithms perform best, in order to draw further conclusions about methods and their combinations in our application context. Our results show trends on the application of specific methods, and indicate that the application of shorter pipelines tends to achieve better results. We also provide empirical evidence suggesting that too intensive dimensionality reduction can have detrimental effects on classifiability, regardless of contamination levels.