Plant-level yield prediction of quinoa in field plots across growth stages using UAV-derived 3D structural and spectral features with machine learning

Quinoa ( Chenopodium quinoa Willd.) is a nutritionally important pseudo-cereal with high environmental adaptability, but its yield potential remains underexploited due to limited precision monitoring tools. Current yield assessment relies on destructive sampling, which limits spatial and temporal resolution. While UAV-based remote sensing is extensively used for cereal yield prediction, quinoa applications remain largely limited to process-based modeling, representing a critical knowledge gap. This study presents the first demonstration of UAV-based, plant-level yield prediction for quinoa using machine learning and 3D structural features. Unlike major cereals, quinoa exhibits highly diverse canopy colors and complex morphology across genotypes and growth stages, making spectral vegetation indices unreliable. Therefore, 3D structural features provide a more robust basis for yield prediction by integrating branching complexity and canopy density, which directly relate to biomass accumulation and reproductive capacity. We conducted UAV surveys combined with machine learning across five growth stages over two seasons (2022 & 2024–2025). RGB and multispectral imagery generated 16 features, including 3D structural parameters and vegetation indices. Five ML regression models were evaluated at each stage, with Random Forest at flowering achieving optimal accuracy (R² = 0.726, RMSE = 16.71 g/m²). Early and late vegetative stages demonstrated slightly lower but still robust prediction accuracy (R² = 0.690–0.704), enabling practical early-season management decisions. Notably, 3D structural features, particularly plant volume and surface area, substantially outperformed spectral indices. These results validate UAV-based ML for non-destructive quinoa yield prediction and demonstrate the essential role of 3D structural features for architecturally complex crops.