Hyperspectral Imaging Technology Combined With Ensemble Learning for Nitrogen Detection in Dendrobium nobile
KUANG Run1, 2, 3, 4, LONG Teng1, 3, 4, LIU Hai-lin2, WU Ji-hui1, 3, 4, LÜ Jin-sheng1, 3, 4, XIE Zi-ran1, 3, 4, LIU Wen-tao1, 3, 4, LAN Yu-bin1, 3, 4, LONG Yong-bing1, 3, 4, WANG Zai-hua2*, ZHAO Jing1, 3, 4*
1. College of Electronic Engineering (College of Artificial Intelligence), South China Agricultural University, Guangzhou 510642, China
2. Environmental Horticulture Research Institute of Guangdong Academy of Agricultural Sciences/Guangdong Provincial Key Lab of Ornamental Plant Germplasm Innovation and Utilization, Guangzhou 510640, China
3. National Center for International Collaboration Research on Precision Agricultural Aviation Pesticides Spraying Technology, Guangzhou 510642, China
4. Guangdong Laboratory for Lingnan Modern Agriculture, Guangzhou 510642, China
Abstract：Dendrobium nobile leaf blades' nitrogen content is a decisive factor for precise fertilization. Traditional nitrogen content detection methods are time-consuming and can completely deplete samples. Efficiently detecting the nitrogen content of D.nobile leaf blades has become a growing concern for herbal medicine cultivation enterprises. To quickly and non-destructively obtain the nitrogen content of D.nobile leaf blades, this study used fresh D.nobile leaf blades as experimental samples. After obtaining their hyperspectral images in the range of 402.6~1 005.5 nm and nitrogen chemical detection values,the images underwent the extraction of regions of interest(ROI), followed by preprocessing of the spectral information within those regions' learner algorithms including Partial Least-Squares Regression (PLSR), Kernel Ridge Regression (KRR), and Support Vector Regression (SVR), as well as ensemble learning algorithms including Random Forest (RF), Bagging, and Adaboost, were utilized to model the nitrogen content of D.nobile. Regression prediction models were constructed based on the full-band spectral information of fresh D. nobile leaf blades and feature bands of spectral information extracted through CARS, and the prediction accuracy was compared. The results showed that when constructing the monitoring model based on the full-band spectral information, the RF model built with spectral data preprocessed by the Savitzky-Golay filtering （SG） method had the best prediction result (R2CV=0.961 4, RMSECV=0.081 8, R2P=0.972 6, RMSEP=0.063 3), and all models achieved R2 values over 0.90. When constructing the regression prediction model based on feature bands extracted through CARS, the Bagging model had the highest accuracy and stability, with the best prediction result observed in the SG-CARS-Bagging model (R2CV=0.938 7, RMSECV=0.100 0, R2P=0.953 5, RMSEP=0.082 6), while the accuracy of the individual learner models KRR and SVR was significantly lower. The CARS algorithm feature extraction removed some important bands, improving modeling efficiency but reducing model accuracy. Therefore, when optimizing the feature parameters of the regression model, it is necessary always to consider the balance between accuracy and efficiency. The research results indicate that ensemble algorithms such as RF, Bagging, and Adaboost have higher stability and prediction accuracy than individual learner algorithms such as PLSR, KRR, and SVR. They are more suitable for analyzing and processing hyperspectral data and have obvious advantages in the nitrogen nutrition monitoring of D.nobile.
[1] JIANG Ying，HUANG Hou-jin(江 滢，黄厚今). Yunnan Journal of Traditional Chinese Medicine and Materia Medica(云南中医中药杂志)，2015，36(1)：77.
[2] LI Ming-yuan，DENG Wei-ze，TANG Xian-fu，et al(李明元，邓维泽，唐显福，等). Journal of Food Science and Biotechnology(食品与生物技术学报)，2014，33(12)：1233.
[3] LINGHU Chu，GU Rong-hui，QIN Li-kang(令狐楚，谷荣辉，秦礼康). Chinese Traditional and Herbal Drugs(中草药)，2021，52(24)：7693.
[4] LIU Hai-lin，LI Jie，YE Guang-ying，et al(刘海林，李 杰，叶广英，等). Southwest China Journal of Agricultural Sciences(西南农业学报)，2022，35(9)：1992.
[5] YU Huan，ZU Chao，YANG Jian-feng，et al(鱼 欢，祖 超，杨建峰，等). Chinese Journal of Tropical Crops(热带作物学报)，2015，36(5)：878.
[6] QIAN Zong-wei，CHEN Hai-li，LIU Ming-chi(潜宗伟，陈海丽，刘明池). Journal of Plant Nutrition and Fertilizer(植物营养与肥料学报)，2011，17(6)：1451.
[7] Gowen A A，O'Donnell C P，Cullen P J，et al. Trends in Food Science & Technology，2007，18(12)：590.
[8] YU Feng-hua，XING Si-min，GUO Zhong-hui，et al(于丰华，邢思敏，郭忠辉，等). Transactions of the Chinese Society of Agricultural Engineering（农业工程学报），2022，38(2)：175.
[9] Abderrahim D，Taoufiq S，Bouchaib I，et al. Chemometrics and Intelligent Laboratory Systems，2023，240：104925.
[10] Jin J，Wu M J，Song G M，et al. Remote Sensing，2022，14(20)：5204.
[11] Zhang M Z，Chen T E，Gu X H，et al. Computers and Electronics in Agriculture，2023，211：108008.
[12] LIU Yuan-yuan，ZHANG Fan，SHI Qi，et al(刘媛媛，张 凡，师 琪，等). Transactions of the Chinese Society of Agricultural Machinery(农业机械学报)，2022，53(6)：295.
[13] ZHANG Jie，XU Bo，FENG Hai-kuan, et al(张 杰，徐 波，冯海宽，等). Spectroscopy and Spectral Analysis(光谱学与光谱分析)，2022，42(6)：1956.
[14] Li H D，Liang Y Z，Xu Q S，et al. Analytica Chimica Acta，2009, 648(1)：77.
[15] Chen K，Peng Y Q，Lu S Y，et al. Chemosphere，2021，274：129802.
[16] Peng J X，Manevski K，Kørup K，et al. Field Crops Research，2021，268：108158.
[17] Muharam F M，Nurulhuda K，Zulkafli Z，et al. Agronomy-Basel，2021，11(5)：915.
