Hyperspectral band selection for food fraud application using self-organizing maps (SOM)

Abstract

Hyperspectral band selection is the process of selecting an optimal set of narrow wavelength bands from a large number over a broad range, typically for one of two purposes: hyperspectral reconstruction or classification. The former seeks to condense the information content of the full resolution spectrum so that the spectrum may be reconstructed from a relatively small subset of wavelength bands. The latter seeks to enable classification based on features contained within this small subset. In this paper, we introduce a new approach for automated band selection based on analysis of the weight planes from a trained self-organizing map. We refer to this approach as the self-organizing map weight plane distance (SOM WPD) method. We evaluate its benefits by using it to select optimal visible/near infrared (VNIR) and fluorescence wavelength bands from a recent fish fraud study where hyperspectral imaging was used to identify the true species of fish fillets. We apply four common machine learning classifiers to perform this species classification and compare the results to those obtained using a genetic algorithm-based method. This latter method optimized band selection for hyperspectral reconstruction, and these same bands were in turn used for classification. The findings presented in this paper show great promise for the SOM WPD which produced higher classification accuracy with two of the four classifiers for the VNIR data and with all four classifiers for the fluorescence data as compared with the genetic algorithm-based method.