Band Selection for Hyperspectral Imaging Using Occlusion-Based Neural Network Optimization
Document Type
Conference Proceeding
Publication Date
2025
Publisher
Society of Photo-optical Instrumentation Engineers (SPIE)
Source Publication
Proceedings of SPIE 13455: Algorithms, Technologies, and Applications for Multispectral and Hyperspectral Imaging XXXI
Source ISSN
9781510686991
Original Item ID
DOI: 10.1117/12.3053261
Abstract
Hyperspectral imaging (HSI) offers detailed spectral information across hundreds of narrow bands, making it a powerful tool for material classification. However, processing all available bands is computationally expensive and impractical for near real-time applications. In this work, an innovative occlusion method is introduced to identify the most informative spectral bands for specified applications, reducing the data dimensionality without compromising classification performance. For a given application, the method requires the simulation of a wide range of environmental conditions by applying various ambient conditions and signal-to-noise ratios (SNRs) to the reflectance spectra of materials of interest. The augmented data sets are then used to train an artificial neural network (ANN), where the HSI bands are inputs to the ANN, to classify materials. The occlusion method is then applied to systematically evaluate and rank the significance of each band by recursively occluding individual bands and assessing the impact on classification performance, thereby enabling the selection of a reduced subset of bands that retains high classification accuracy. As a case study, this method is applied to the detection of non-visible bruised areas in apples. We simulate HSI images consistent with a Resonon Pika IR+ camera containing 336 bands and show that our approach can reduce the data to six bands while achieving 98.3% HSI classification accuracy and 0.004% segmentation error, demonstrating its efficiency for application-specific band selection. This enables faster and more practical HSI deployment, including integration into smart cameras for edge computing and real-time analysis.
Recommended Citation
Zhuang, Mutian; Ratliff, Bradley M.; and Hayat, Majeed M., "Band Selection for Hyperspectral Imaging Using Occlusion-Based Neural Network Optimization" (2025). Electrical and Computer Engineering Faculty Research and Publications. 823.
https://epublications.marquette.edu/electric_fac/823
Comments
Published as part of the Proceedings of SPIE 13455, Algorithms, Technologies, and Applications for Multispectral and Hyperspectral Imaging XXXI, 2025. DOI.