Proximal hyperspectral detection of rice and weed: characterization and discriminant analysis

Zhentao wang¹Tenghui Lin²huijie li¹Yanling Yin³YuTing Suo⁴Fengjie Cai¹Yulin Li^1*

• ¹Shihezi University College of Mechanical and Electrical Engineering, Shihezi, China
• ²Northeast Agricultural University, Harbin, China
• ³Shihezi University School of Medicine, Shihezi, China
• ⁴Yantai Agricultural Technology Popularization Center, yantai, China

Weeds represent a critical component of agricultural biodiversity and contribute to a range of ecosystem services, yet they remain a major constraint on global crop production. Remote sensing technology, particularly hyperspectral imaging, has advanced from spectral response patterns to species identification and vegetation monitoring. Consequently, the ability to accurately map weed species and assess their physiological activity in agricultural settings is of growing important. In this study, we established a hyperspectral library of rice and weed species in cold regions of northern China, comprising a total of 36 species. Using a ground-based hyperspectral camera (SPECIM-IQ), we collected 1080 hyperspectral images and extracted representative spectral reflectance curves for rice and 35 weed species. We employed canopy spectral profile characteristics, vegetation indices, and principal component analysis (PCA) to characterize and explain the differences among various weeds. A novel deep learning network, SS-CNN, was developed to identify rice and weed species from hyperspectral imagery, and ablation experiments were conducted to evaluate its performance. When the training sample size (Tr) was set at 70%, the SS-CNN model outperformed the comparative models with the best identification results (overall accuracy (OA): 99.910%, average accuracy (AA): 99.502%, Kappa: 0.9991). Even at a reduced training sample size of 5%, the SS-CNN algorithm maintained optimal classification performance (OA: 95.370%; AA: 86.468%; Kappa: 0.9518). This study demonstrates the application of proximal hyperspectral remote sensing and deep learning networks for rice and weed identification and characterization in harsh field scenarios. It provides a valuable baseline for understanding the hyperspectral characteristics of paddy field weed stress and monitoring their growth status.