Estimation of relative chlorophyll content in winter wheat employing hyperspectral reflectance: A comprehensive analysis of data-driven ensemble learning methods

Ensemble learning is increasingly used for remote sensing-based plant phenotyping data. A thorough evaluation of its effectiveness in estimating relative chlorophyll content is essential to optimize model selection and enhance prediction accuracy. This study aimed to assess the performance of parallel, sequential, and hybrid ensemble learning approaches, as well as individual machine learning models, for cross-environment estimation of SPAD-based chlorophyll content using hyperspectral reflectance data. Canopy hyperspectral reflectance and SPAD measurement were collected from winter wheat during the late growth stages across two distinct environments. Parallel and sequential ensemble learning strategies were represented by random forests (RF) and eXtreme gradient boosting (XGBoost), respectively. The hybrid ensemble integrated the performance of K-nearest neighbors, support vector machine, partial least squares regression, generalized linear model (GLM), deep neural network, and Gaussian process. Gray relational analysis (GRA) was employed to evaluate band features, and the positive association between feature quality and prediction accuracy validated its effectiveness. During modeling, XGBoost (R 2 = 0.657–0.658) underperformed compared to RF (R 2 = 0.687–0.691). GLM consistently excelled across most hybrid ensemble members in most feature intervals (R 2 = 0.137–0.692, 0.512–0.723), achieving superior prediction accuracy across several feature intervals and also outperforming RF (R 2 = 0.156–0.687, 0.535–0.691). Among the six forecast combination strategies used in the hybrid ensemble, inverse rank (R 2 = 0.640–0.726) demonstrated robust performance across different feature intervals but provided only marginal improvement over the best individual ensemble member. Moreover, incorporating RF and XGBoost in the hybrid ensemble did not result in significant accuracy gains. To balance computational efficiency and prediction accuracy, GRA-based optimal features combined with GLM modeling are recommended for similar applications, rather than relying on complex ensemble learning methods. These findings offer valuable insights for estimating relative chlorophyll content and hold potential for large-scale estimation of crop traits.