Rapid and non-destructive monitoring of SPAD values in fresh tea leaves is a key technical approach for achieving precision nitrogen management in tea plantations. Given that the high moisture content and substantial inter-varietal heterogeneity of fresh tea leaves lead to insufficient generalization capacity of existing spectral monitoring models, this study aimed to identify an optimal technical pipeline of “spectral preprocessing-feature selection-modeling” tailored to the characteristics of fresh tea leaves. A total of 2 844 fresh tea leaf samples were collected during the spring tea harvest seasons of 2025 and 2026. After excluding two samples whose SPAD readings exceeded the coefficient of variation threshold, 2 842 pairs of valid data were retained. Spectral data were acquired using an ASD FieldSpec 4 spectroradiometer, and SPAD values were simultaneously measured with a SPAD-502 Plus chlorophyll meter. The effects of the combinations of five spectral preprocessing methods with three feature selection strategies [competitive adaptive reweighted sampling (CARS), successive projections algorithm (SPA), and uninformative variable elimination (UVE)] on the predictive performance of four modeling algorithms-partial least squares regression (PLSR), random forest (RF), support vector regression (SVR), and back propagation neural network (BPNN)-were systematically compared. Model hyperparameters were optimized via grid search combined with cross-validation. The results showed that the combination of first-derivative preprocessing and the UVE algorithm yielded superior performance, selecting 225 key spectral bands from the full spectrum. The RF model built on this feature subset achieved high accuracy on the training set (R2=0.955 0, RMSE=2.366 5) but exhibited a marked decline in performance on the test set (R2=0.655 4, RMSE=6.414 7), indicating limited generalization capacity under cross-variety and cross-harvest-period conditions. After hyperparameter optimization, the independent validation performance of the RF model was improved (R2=0.703 2, RMSE=6.022 4), effectively alleviating the overfitting problem. Cross-year validation further indicated that the inter-annual generalization capacity of the model still has room for improvement. In conclusion, the optimized RF model established using the 225 feature bands selected by first-derivative preprocessing combined with UVE effectively reduces data dimensionality while preserving chlorophyll-sensitive information, providing methodological support for the rapid and non-destructive prediction of SPAD values in fresh tea leaves across different varieties and harvest periods, as well as for precision nitrogen management in tea plantations.
经过特征选择后,利用筛选后的特征变量构建茶鲜叶SPAD值的估测模型。模型评价选择决定系数R2和均方根误差(Root mean square error,RMSE)为评估指标。R2取值范围为0~1,越接近1代表回归模型预测精度越高。RMSE与SPAD值单位相同,其值越小,表明回归模型预测误差越小。各指标表达式分别如下:
BatoolD, ShahbazM, Shahzad AsifH, ShaukatK, AlamT M, HameedI A, RamzanZ, WaheedA, AljuaidH, LuoS H. A hybrid approach to tea crop yield prediction using simulation models and machine learning[J]. Plants, 2022, 11(15): 1925
[2]
WuW D, PeiG Q, LuZ Q, ZhouB, QianX Y, WangB J, YangL N. Lightweight multi-view fusion network for non-destructive chlorophyll and nitrogen content estimation in tea leaves using front and back RGB images [J]. Agronomy, 2025, 15(10): 2355
TuS P, HuangH, DuQ, TaoL J. Comparison of photosynthetic characteristics and chlorophy Ⅱ fluorescence parameters of different Camellia sinensis cultivars [J]. Acta Agriculturae Universitatis Jiangxiensis, 2021, 43(5): 1098-1106 (in Chinese)
[5]
ChenJ M, WuS H, DongF, LiJ L, ZengL T, TangJ C, GuD C. Mechanism underlying the shading-induced chlorophyll accumulation in tea leaves[J]. Frontiers in Plant Science, 2021, 12: 779819
[6]
RazaA, LiL, LuoL, AsgharM A, ShoaibN, YinC Y. Rational application of combined fertilizers improved tea growth and quality components[J]. Journal of the Science of Food and Agriculture, 2025, 105(7): 3923-3930
LiuY D, LinX D, GaoH G, WangS, GaoX. Research on tea Cephaleuros virescens kunze model based on chlorophyll fluorescence spectroscopy[J]. Spectroscopy and Spectral Analysis, 2021, 41(7): 2129-2134 (in Chinese)
ZouZ H, ShenC, LiX, ZhangL P, ZhangL, HanW Y, YanP. Status quo of nitrogen fertilizer application and loss in tea garden of China[J]. Journal of Plant Nutrition and Fertilizers, 2021, 27(1):153-160 (in Chinese)
PanJ H, ChengR G, WangP, YuW S. Hyperspectral nitrogen retrieval of tea plant canopy by derivative transformation combined with machine learning[J]. Jiangsu Agricultural Sciences, 2026, 54(4): 98-105 (in Chinese)
[15]
VaddiR, Phaneendra KumarB L N, ManoharanP, AgilandeeswariL, SangeethaV. Strategies for dimensionality reduction in hyperspectral remote sensing: A comprehensive overview[J]. The Egyptian Journal of Remote Sensing and Space Sciences, 2024, 27(1): 82-92
[16]
WangS P, AltanerC, FengL, LiuP P, SongZ Y, LiL Q, GuiA H, WangX P, NingJ M, ZhengP C. A review: Integration of NIRS and chemometric methods for tea quality control-principles, spectral preprocessing methods, machine learning algorithms, research progress, and future directions [J]. Food Research International, 2025, 205:115870
ZhaoX H, FengX Q, NieX J, ZhangY, GuoE H. Hyperspectral estimation of soil phosphorus content in mining area based on CARS-CNN[J]. Chinese Journal of Ecology, 2024, 43(12): 3553-3562 (in Chinese)
MengS, LiX G, JiangY D, EZIZMamattursun. Estimation of soil organic carbon content in lakeside oases based on CWT-RF model[J]. Journal of China Agricultural University, 2023, 28(3): 198-207 (in Chinese)
JiangJ, ZhaoZ W, CaiC, ZhangJ S, ChengZ Q. Hyperspectral estimation of tea leaves water content under the influence of dust retention[J]. Spectroscopy and Spectral Analysis, 2021, 41(11): 3532-3537 (in Chinese)
LiX L, ZhangD Y, DongY L, JinJ J, HeY. Spectral rapid detection of phytochemicals in tea( Camellia sinensis ) based on convolutional neural network [J]. Journal of China Agricultural University, 2021, 26(11): 113-122 (in Chinese)
[27]
JiangJ, JiH T, YanY, ZhaoL Y, PanR Y, LiuX J, YinJ, DuanY, MaY C, ZhuX J, FangW P. Mining sensitive hyperspectral feature to non-destructively monitor biomass and nitrogen accumulation status of tea plant throughout the whole year[J]. Computers and Electronics in Agriculture, 2024, 225: 109358
LiaoJ, FangH S, SuY J, WangY J, ZhangY L, SunY L, FangY G. Change law of water loss in fresh tea leaves under different spreading environments[J]. Chinese Agricultural Science Bulletin, 2023, 39(4): 160-164 (in Chinese)
[30]
WeiY Z, WuF Y, XuJ, ShaJ J, ZhaoZ F, HeY, LiX L. Visual detection of the moisture content of tea leaves with hyperspectral imaging technology[J]. Journal of Food Engineering, 2019, 248: 89-96
ZhuL Q, QinD M, ZhaoL J, DengB S, LiuS Y, JiangC J.Leaf epidermal micromorphological features and their systematic significance of six wild species of Camellia chrysantha[J]. Bulletin of Botanical Research,2021, 41(6): 841-850 (in Chinese)
QinA Y, ChenE H, LaoF Z, ChenD H, ZhongY, MoG H. Leaf anatomical determinants of stress resistance and productivity: A comprehensive evaluation of 50 tea accessions in Guangxi[J]. Chinese Journal of Tropical Crops, 2025, 46(10): 2376-2389 (in Chinese)
LiuD N, GongX J, LiL Y, HuangF, YaoY, XuY Q, GaoY, LuoF. Analysis of photosynthetic and fluorescence characteristics of albino tea plants[J]. Journal of Tea Science, 2023, 43(6): 757-768 (in Chinese)
[37]
ChenY J, HanY X, HeS Q, ChengQ H, TongH R. Differential metabolic profiles of pigment constituents affecting leaf color in different albino tea resources[J]. Food Chemistry, 2025, 467: 142290
WangX L, HanL J, YangZ L, XiaoW H, WangJ Y, YuL Y, DaiJ W, ZhuL Q. Influence of preprocessing methods on the rapid quantitative models for crop straw nutrients using different near-infrared spectroscopy devices data[J]. Transactions of the Chinese Society of Agricultural Engineering, 2025, 41(9): 216-226 (in Chinese)
DiwuP Y, BianX H, WangZ F, LiuW. Study on the selection of spectral preprocessing methods[J]. Spectroscopy and Spectral Analysis, 2019, 39(9): 2800-2806 (in Chinese)
[44]
YuanR R, LiuG S, HeJ G, WanG L, FanN Y, LiY, SunY R. Classification of Lingwu long jujube internal bruise over time based on visible near-infrared hyperspectral imaging combined with partial least squares-discriminant analysis [J]. Computers and Electronics in Agriculture, 2021, 182: 106043
[45]
ZhouL, WuH B, JingT T, LiT H, LiJ S, KongL J, ZhouL N. Estimation of relative chlorophyll content in lettuce ( Lactuca sativa L. ) leaves under cadmium stress using visible: Near-infrared reflectance and machine-learning models [J]. Agronomy, 2024, 14(3): 427
[46]
DuL, JinZ L, ChenB W, ChenB W, GaoW, YangJ, ShiS, SongS L, WangM M, GongW, WangW. Application of hyperspectral LiDAR on 3-D chlorophyll-nitrogen mapping of rohdea japonica in laboratory [J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021, 14: 9667-9679
[47]
GaoL, DarvishzadehR, SomersB, JohnsonB A, WangY, VerrelstJ, WangX F, AtzbergerC. Hyperspectral response of agronomic variables to background optical variability: Results of a numerical experiment[J]. Agricultural and Forest Meteorology, 2022, 326: 109178
[48]
Jarro-EspinalI, Huanuqueño-MurilloJ, Quille-MamaniJ, Quispe-TitoD, Ramos-FernándezL, Pino-VargasE, Torres-RuaA. Field-scale rice yield prediction in northern coastal region of Peru using sentinel-2 vegetation indices and machine learning models [J]. Agriculture, 2025, 15(19): 2054
ZhouJ, LeiG, LuX, DengM, KanF, HuangY. Research on the comparison of remote sensing extraction methods for planting structures in medium and high resolution images[J]. Water Saving Irrigation, 2026: 1-15 (in Chinese)
[51]
SunZ G, WangG T, LiP F, WangH, ZhangM, LiangX W. An improved random forest based on the classification accuracy and correlation measurement of decision trees[J]. Expert Systems with Applications, 2024, 237: 121549
XiongX Y, YangX Z, ZhaoY C, LiW P. Remote sensing inversion of forest biomass based on hyperparametric optimized random forests algorithm[J]. Journal of Central South University of Forestry & Technology, 2024, 44(5): 102-111 (in Chinese)