An End-to-End Framework for the Classification of Hyperspectral Images in the Wood Domain

Hyperspectral images consist of a multitude of spectral bands for each pixel. Spectral bands provide information about wavelengths that may cover a larger spectrum of what the human eye may see. In the hyperspectral domain, the classification of hyperspectral images is usually addressed by taking into account only the spectral information. However, in the wood domain, spatial information is also relevant. To bridge this gap, this paper proposes a CNN-based end-to-end framework for the classification of hyperspectral images in the wood domain. The proposed framework consists of a spatial and spectral classifier that are integrated to make the final prediction. Each classifier is built by adapting a general image classifier, which is suitable for the classification of three-band images, to handle hyperspectral images. The framework is trained and validated on a real dataset, provided by a company working in the wood domain to detect wood fungi. The results obtained have shown that the proposed framework is a lightweight and effective approach for the recognition of wood fungi categories. The framework outperforms a benchmark classifier by 17% and can generate a classification map of hyperspectral images of wood boards of any size with an accuracy of 96%.