Target Detection Based on Spectral Derivation in HSI Shadow Region Classified by Convolutional Neural Networks

Abstract

Because of the low reflection value of the shadow regions in hyperspectral image (HSI), these regions were deleted directly or ignored in target detection or classification. There are some studies on improving the reflectivity of shadow regions, but it is still difficult to determine the actual substances contained in shadow regions. In this paper, an improved target detection method in shadow regions in HSI is proposed by combining the convolution neural network (CNN) and the adaptive cosine consistency estimation (ACE) of the spectral derivative image. Firstly, the derivative image could be obtained by deriving the original HSI in the spectral dimension. The following steps would be performed simultaneously on the original HSI and the derivative image. Next, the shadow regions in HSI could be determined through two-dimensional (2D) CNN model whose main parameters have been adjusted to optimize the network performance. Finally, the substances contained in the shadow regions would be detected by ACE algorithm. The performance of the proposed method was evaluated and analyzed on the real-world HSIs. The experiments show that the spectral derivation can help to improve the target detection and it is worth taking into account the shadow regions for processing HSI data.