Knowledge Guided Classification Of Airborne Hyperspectral Images With Deep Convolutional Neural Network

Hyperspectral Images (HSI) contains hundreds of spectral information, which provides detailed spectral information, has an inherent advantage in land cover classification. Benefiting from the previous studies on hyperspectral mechanisms, hyperspectral technology has achieved significant progress in classification. Deep learning technology, with remarkable learning ability, can better extract the spatial and spectral information of HIS, which is essential for classification. However, the research and application of deep learning in HIS classification are still insufficient, especially in terms of combining with prior knowledge, which has an advantage in data optimization. In this paper, a novel CNN network, name IUNet, is proposed for airborne hyperspectral classification. Besides, Besides, a series of knowledge-guided methods such as Radiation Consistency Correction (RCC) and Minimum Noise Fraction (MNF) were introduced to optimize the HIS data. Selected spectral indexes are employed to improve the classification accuracy according to the characteristics of the target. The HyMap images from Gongzhuling area of Jilin Province are used for experiments, and the experimental results show that the application of prior knowledge in data optimization can significantly improve the classification performance of hyperspectral classification based on deep learning.