Peanut origin identification: A hyperspectral system combined with a convolution transformer hybrid dense network

Abstract

The quality of peanuts from different origins varies widely due to the natural environment and growing methods. It is common for peanuts from low-quality origins to be passed off as high-quality. In this work, we propose an effective spectral information processing method and combine it with a fast and non-destructive hyperspectral detection technique to identify peanuts from different origins accurately. First, we use a hyperspectral system to obtain the spectral information of peanuts. Second, this paper proposes a Convolution Transformer Hybrid Block (CTHB) to extract spectral information from local and global aspects and to fuse the extracted feature information adaptively. Finally, combined with the data characteristics of spectral information, we introduce a Residual Dense Connection to avoid feature degradation effectively, and we tightly connect multiple CTHBs using the Residual Dense Connection. This forms a Convolution Transformer Hybrid Dense Network (CTHD-Net) to effectively recognize peanuts’ spectral information. In comparing multi-model results, the CTHD-Net shows the best classification performance, reaching 98.03 % accuracy, 98.07 % precision, and 97.55 % recall rate. The results show that the combination of the CTHD-Net and hyperspectral system effectively distinguishes the quality of peanuts from different producing areas. This provides an effective technical way for agriculture and post-harvest quality inspection.