Fusing UAV multiple data and phenology to predict crop biomass

• Machine learning models have varying preferences for feature types. • Stacked ensemble learning approach yielded the more accurate AGB prediction. • Multimodal data fusion improved the transferability of the AGB prediction models. • Analyzed the importance of UAV features across wheat development stages. Robust quantification of crop status in real-time is essential for agile decision-making. While use of unmanned aerial vehicle data (UAV) appears promising in this vein, the contribution and transferability of various features (e.g. vegetation indices, plant height and texture features) in crop above-ground biomass (AGB) prediction remain poorly understood. Here, our objectives were to (1) evaluate the performance of various machine learning (ML) algorithms in the synthesis of multiple features, (2) elicit the contribution of various UAV features, (3) assess the transferability of features across growth stages and sites. Four field experiments, incorporating several water and nitrogen treatments across two sites, were assembled for use in AGB prognostics. We invoked four ML algorithms—Random forest (RF), Lasso regression (LR), K-nearest neighbors (KNN) and a stacked ensemble integrating the three methods (SML)—to predict wheat AGB using multiple UAV data and phenological information. Additionally, interpretable ML techniques were employed to elucidate the influence of UAV features on AGB prediction across growth stages. Our results showed that all algorithms exhibited robust performance in predicting wheat biomass, with RMSE values of 1.64, 1.71, 1.71, and 1.57 Mg ha −1 for RF, LR, KNN, and SML, respectively. RF predominantly relied on plant height features, LR leveraged vegetation indices, and KNN prioritized texture features, while SML synthesized the advantages of multiple ML algorithms. Fusion of multiple datasets amplified model prognostic capacity and scalability, with R 2 and rRMSE of 0.92 and 22 % when using data from external sites. Features pertaining to vegetation indices and plant height during vegetative growth and around flowering had seminal contributions of model predictions. Texture features significantly reduced the saturation effect during the reproductive stage but diminished the model’s transferability during the vegetative stage. Complementarity among data types enhanced effectiveness of ensemble machine learning, which leverages strengths of diverse data to improve the accuracy and robustness of AGB predictions. Future studies could combine multiple sources of remote sensing, such as LiDAR and thermal infrared alongside system modeling, to improve ML accuracy and generalization capability.