Early estimation of faba bean yield based on unmanned aerial systems hyperspectral images and stacking ensemble

Abstract

Abstract Faba bean is a vital legume crop, and its early yield estimation can improve field management practices. In this study, unmanned aerial system (UAS) hyperspectral imagery was used for the first time to estimate faba bean yield early. Different basic algorithms, including random forest (RF), support vector machine (SVM), k-nearest neighbor (KNN), partial least squares regression (PLS), and eXtreme Gradient Boosting (XGB), were employed along with stacking ensemble learning to construct the faba bean yield model and investigate factors influencing model accuracy. The results are as follows: when using the same algorithm and growth period, integrating texture information into the model improved the estimation accuracy compared to using spectral information alone. Among the base models, the XGB model performed the best in the context of growth period consistency. Moreover, the stacking ensemble significantly improved model accuracy, yielding satisfactory results, with the highest model accuracy (R2) reaching 0.76. Model accuracy varied significantly for models based on different growth periods using the same algorithm. The accuracy of the model gradually improved during a single growth period, but the rate of improvement decreased over time. Data fusion of growth period data helped enhance model accuracy in most cases. In conclusion, combining UAS-based hyperspectral data with ensemble learning for early yield estimation of faba beans is feasible, therefore, this study would offer a novel approach to predict faba bean yield.