文献类型： 外文期刊

作者： Shen, Jian ¹ ; Huang, Yurong ² ; Chen, Wenqian ² ; Li, Mengjun ¹ ; Tan, Wei ² ; Wang, Ronghui ¹ ; Deng, Yujia ² ; Gong, Yingting ¹ ; Ai, Shaoying ¹ ; Liu, Nanfeng ² ;

作者机构： 1.Guangdong Acad Agr Sci, Inst Agr Resources & Environm, Key Lab Plant Nutr & Fertilizer South Reg, Natl Agr Expt Stn Soil Qual,Minist Agr & Rural Aff, Guangzhou 510640, Peoples R China

2.Sun Yat Sen Univ, Sch Geog & Planning, Guangdong Prov Key Lab Urbanizat & Geosimulat, Carbon Water Res Stn Karst Reg Northern, Guangzhou 510275, Peoples R China

关键词： model transferability; foliar nutrient diagnosis; hyperspectral remote sensing; machine learning; PROSPECT-PRO

期刊名称：REMOTE SENSING （ 影响因子：4.1； 五年影响因子：4.8 ）

ISSN：

年卷期： 2025 年 17 卷 4 期

页码：

收录情况： SCI

摘要： Fresh sweet and waxy maize (Zea mays) are valuable specialty crops in southern China. Hyperspectral remote sensing offers a powerful tool for detecting maize foliar nutrients non-destructively. This study aims to investigate the capability of leaf spectroscopy (SVC HR-1024i spectrometer, wavelength range: 400-2500 nm) to retrieve maize foliar nutrients. Specifically, we (1) explored the effects of nitrogen application rates (0, 150, 225, 300, and 450 kgNha-1), maize cultivars (GLT-27 and TGN-932), and growth stages (third leaf (vegetation V3), stem elongation stage (vegetation V6), silking stage (reproductive R2), and milk stage (reproductive R3)) on foliar nutrients (nitrogen, phosphorus, and carbon) and leaf spectra; (2) evaluated the transferability of the regression and physical models in retrieving foliar nutrients across maize cultivars. We found that the PLSR (partial least squares regression), SVR (support vector machine regression), and RFR (random forest regression) regression model accuracies were fair within a specific cultivar, with the highest R2 of 0.60 and the lowest NRMSE (normalized RMSE = RMSE/(Max - Min)) of 17% for nitrogen, R2 of 0.19 and NRMSE of 21% for phosphorous, and R2 of 0.45 and NRMSE of 19% for carbon. However, when these cultivar-specific models were used to predict foliar nitrogen across cultivars, lower R2 and higher NRMSE values were observed. For the physical model, which does not rely on the dataset, the R2 and NRMSE for foliar chlorophyll-a and -b (Cab), carotenoid (Cxc), and equivalent water thickness (EWT) were 0.76 and 15%, 0.67 and 34%, and 0.47 and 21%, respectively. However, the prediction accuracy for foliar nitrogen, expressed as foliar protein in PROSPECT-PRO, was lower, with an R2 of 0.22 and NRMSE of 27%, which was comparable to that of the regression models. The primary reasons for this limited transferability were attributed to (1) the insufficient number of samples and (2) the lack of strong absorption features for foliar nutrients within the 400-2500 nm wavelength range and the confounding effects of other foliar biochemicals with strong absorption features. Future efforts are needed to investigate the physical mechanisms underlying hyperspectral remote sensing of foliar nutrients and incorporate transfer learning techniques into foliar nutrient models.