Spatio-temporal prediction of total and legume dry matter yield using UAV-borne RGB and multispectral images in alfalfa-grass mixtures

Abstract Accurate and timely forage yield prediction in alfalfa-grass mixtures (AGM) is essential for supporting precision agriculture management decisions. This study aimed to develop and evaluate UAV-borne remote sensing models to predict total dry matter yield (DMY) and legume dry matter yield (LY) across multiple harvests and field sites. UAV-borne high-resolution true-colour images were used to derive canopy height models via structure-from-motion. At the same time, multispectral imagery enabled the calculation of reflectance-based vegetation indices. Biomass was destructively sampled, and DMY and LY were determined through drying and botanical fractioning. A total of 276 biomass samples were collected over four harvests, including samples from three AGM fields. To predict DMY and LY, two machine learning regression models (random forest and extreme gradient boosting) were trained and validated using leave-spatial-temporal-group-out cross-validation to ensure robustness across locations and time. Random forest models using fused spectral and height data achieved the best performance, with median prediction errors of 0.51 t ha⁻¹ for DMY (median R² = 0.49) and 0.40 t ha⁻¹ for LY (median R² = 0.65), demonstrating good generalizability under varying agronomic conditions. The study highlights the potential of combining UAV-borne height and spectral data for high-resolution yield mapping in complex forage systems. Predictive maps of DMY and LY provide spatial insights that can inform management and support sustainable nitrogen cycling in crop rotations.