UAV remote sensing and ensemble learning for predicting rice aboveground biomass: how to select base and meta learners?

Aboveground biomass (AGB) is a practical agroecological indicator, particularly suitable for monitoring crop growth, predicting yield, and evaluating crop–environment–management interactions. In most cases, AGB prediction relies on destructive sampling. Although accurate, this approach is time-consuming and labor-intensive and has limited spatial representativeness, making it difficult to apply in large-scale, high-frequency dynamic monitoring scenarios. With high spatiotemporal resolution, operational flexibility, and cost advantages, unmanned aerial vehicle (UAV) remote sensing has become an ideal tool for continuous, large-scale assessment of crop aboveground biomass. The combination of UAV remote sensing and machine learning (ML) has significantly advanced field-scale AGB prediction research. However, in heterogeneous field environments, single machine learning algorithms often lack prediction robustness due to model structure and parameter uncertainty. Ensemble learning frameworks (ELFs) integrate the predictions of multiple base learners through a meta-learner, leveraging model complementarity to improve prediction accuracy and stability. Nevertheless, systematic evaluations of meta-learner selection and the effects of the number and diversity of base learners remain scarce.

Recently, the team led by Xinwei Li at the College of Resources and Environment, Anhui Science and Engineering University, and collaborators published a research paper titled “Remote sensing using unmanned aerial vehicles and ensemble learning for rice aboveground biomass prediction: insights into selecting base and meta learners” in Engineering Agriculture, an Engineering cluster journal of the Chinese Academy of Engineering (DOI: 10.15302/J-FASE-2026684). Using vegetation indices (VIs) and texture features (TFs) extracted from multi-temporal UAV images, the study constructed an ensemble learning framework integrating nine machine learning algorithms and three meta-learners: linear model (LM), random forest, and Bayesian model averaging (BMA), and systematically evaluated the effects of base learner configurations on prediction accuracy.

The field experiment was conducted from June to October 2020 in Xiaogang Village, Fengyang County, Anhui Province. A split-plot design was adopted, with nitrogen application rate as the main-plot factor and rice cultivar as the subplot factor, totaling 36 experimental plots. A DJI Phantom 4 Multispectral UAV was used to acquire multispectral images at six key growth stages of rice, and destructive sampling was conducted simultaneously to obtain measured AGB values. Multiple VIs and gray-level co-occurrence matrix (GLCM)-based TFs were extracted from the images. Recursive feature elimination was used for sensitive feature selection, and seven VIs and eight TFs were ultimately selected. The training set and independent test set were constructed at a 7:3 ratio. Five-fold cross-validation was used for base learner training, and the entire workflow was repeated for 20 independent iterations to eliminate the influence of randomness.

The results showed that models using the fused feature set of VIs and TFs were significantly better than models relying on a single feature set. Among the nine base learners, Gaussian process regression (GPR) performed best. On the VIs + TFs test set, GPR achieved the highest R² = 0.769 and the lowest RMSE = 1.83 t·ha⁻¹; random forest regression (RFR) ranked second (R² = 0.763, RMSE = 1.85 t·ha⁻¹). All three meta-learners outperformed the best single base learner. Among them, the BMA meta-learner achieved the highest accuracy on the fused feature set (R² = 0.795, RMSE = 1.73 t·ha⁻¹), with the best generalization ability and stability tied for the highest with the LM meta-learner. Meta-learner performance was highly sensitive to the composition of the base learner pool: when the number of base learners was five, accuracy and stability reached an optimal balance; the hybrid ensemble composed of linear kernel models (SMR, PLSR, LASSO, and RR) and GPR achieved the highest accuracy (R² = 0.795, RMSE = 1.73 t·ha⁻¹). Ensembles containing GPR were significantly better than combinations without GPR, whereas including the weakest base learner, PLSR, reduced overall performance.

This study demonstrates that meta-learner selection and base learner composition have a critical influence on the prediction accuracy of ELFs. The hybrid ensemble strategy combining a BMA meta-learner with linear kernel models and GPR provides clear methodological guidance for high-precision UAV-based crop AGB monitoring. These findings have practical application value for crop growth dynamics monitoring and yield prediction in precision agriculture.