Estimating grass-clover mixtures quality using field spectroradiometry and stacking-based ensemble learning models

In this study, we designed a stacking ensemble machine learning method for accurate estimation of grass-clover biomass quality from high-resolution hyperspectral field data. Basic learners were support vector machine (SVM), K-Nearest Neighbor (KNN), random forest (RF), decision tree (DT), partial least square regress (PLS), shallow artificial- (ANN) and deep multi-layer perceptron (MLPs) neural networks. Three meta learners in the second layer were RF (stackingRF), eXtreme Gradient Boosting (stackingXGB), and CatBoost (stackingCat). We used 6 different datasets for the estimation, including the raw spectral reflectance, its 1 st and 2 nd derivatives, and their Standard Normal Variate (SNV). Quality indices were crude ash, crude protein, crude fat, crude fiber, neutral detergent fiber (NDF), acid detergent fiber (ADF), water-soluble carbohydrates (WSC), sugar, and fructans. The results showed that spectral feature selection by the PLS was useful in preprocessing. All machine learning methods predicted well quality from the raw reflectance, with PLS performing best (R 2 = 0.64–0.87) for all indices except crude ash. The stacked ensemble learning benefited the predictions, especially with stackingRF and stackingCat as meta learners (R2 increased by 2–8 % compared with the average of the four best machine learning algorithms in the first layer). SNV generally performed better than the original reflectance across most datasets. In particular, SNV applied to the 2nd derivative, combined with either PLS or StackingRF, yielded the best predictions for carbohydrates (sugar, fructans, WSC) and fiber (crude fiber, NDF, ADF), achieving the highest R 2 values (0.74–0.87) and the lowest nRMSE (5–15 %) among all pre‑processing options. The red‑edge region (700–780 nm) is the most important for all quality indices, except for ADF. This is among the first studies to systematically evaluate stacking ensemble learning, using the outputs of multiple machine‑learning algorithms as inputs, for predicting biomass quality and supporting field management in complex grass–clover mixtures. PLS‑based feature selection combined with stacking ensemble learning on second‑derivative SNV spectra appears particularly promising and warrants testing at larger spatial and temporal scales.