Hyperspectral proximal sensing of leaf chlorophyll content of spring maize based on a hybrid of physically based modelling and ensemble stacking

• Published in: Computers and electronics in agriculture 2023-05, Vol.208, p.107745, Article 107745
• Main Authors: Huang, Xi, Guan, Huade, Bo, Liyuan, Xu, Zunqiu, Mao, Xiaomin
• Format: Article
• Language: English
• Subjects: Hybrid inversion; Hyperspectral data; Leaf chlorophyll content; Machine learning; PRAOSAIL; Stacking regression
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2023.107745

•Leaf chlorophyll content (LCC) and canopy hyperspectral reflectance were measured for maize.•Ensemble stacking with bands selection, base regression algorithms and PROSAIL was used to predict LCC.•Back propagation neural network was the best performing model among base models.•Wavelengths at 698, 702 and 705 nm were most sensitive for LCC.•Ensemble stacking models performed substantially better than base models. Leaf chlorophyll content (LCC) is an important indicator for evaluating crop nutritional status and environmental stress. For the purpose of achieving rapid, non-destructive, and real-time LCC monitoring, experiments of spring maize under three drip irrigation levels and three film mulching conditions were conducted in the Shiyang River Basin of Northwest China in 2020 and 2021. We measured LCC and canopy hyperspectral reflectance during the maize growth period. Several two-layer ensemble stacking models were built to predict LCC from field hyperspectral measurement. The layer-1 base regression algorithms (i.e., support vector regression (SVR), back propagation neural network (BPNN), partial least square regression (PLSR)) were trained with synthetic data from the physically-based PROSAIL modelling. Multiple linear regression (MLR) was adopted for the layer-2 meta model, which was trained with the layer-1 modelling output and field observed LCC. The effects of the selected sensitive hyperspectral bands were also examined via three methods: recursive feature elimination (RFE), correlation analysis (CA) and variable importance in the projection analysis (VIP). Finally, the predictability and robustness of 16 models, composed of four types of input variables and four regression algorithms (three base regression algorithms and ensemble stacking algorithm), were tested using three field measured datasets. The results showed that the wavelengths at 698, 705, 693, 695, 697, 699, 700, 703, 709 and 712 nm were most sensitive for LCC. In the base regression models, the BPNN was the best performing model, followed by PLSR and SVR. The ensemble stacking models (R2: 0.78–0.91, RMSE: 3.27–9.34 μg cm−2, NRMSE: 7–20%) performed substantially better than the base models (R2: 0.28–0.68, RMSE: 10.53–30.83 μg cm−2 and NRMSE: 22–65%) did. No statistically significant difference was found in the stacking model performance between the two field years. We highly recommend using PROSAIL modelling to train machine-learning based stacking models for the proximal sensing of