Fusion of UAV-based 3D mesh and spectral features improves quinoa biomass and LAI estimation across genotypic and temporal variations

Le résumé fourni par la source

• First study integrating UAV-based 3D mesh models with spectral indices for quinoa monitoring. • 3D mesh metrics proved highly suitable for quinoa due to its unique architectural features. • Plant volume and RGB COM models effectively estimated biomass and leaf area index. • Fusion models achieved high accuracy and remained robust against genetic and temporal changes. The increasing global importance of quinoa necessitates the development of advanced monitoring techniques for crop management and breeding. The potential of unmanned aerial vehicle (UAV)-based remote sensing for monitoring quinoa growth parameters remains largely unexplored, particularly regarding aboveground biomass (AGB) monitoring and comprehensive analysis integrating 3D structural information with spectral indices for improved estimation of AGB and leaf area index (LAI) across multiple growth stages. This study used UAV-derived 3D mesh models combined with multispectral and RGB vegetation indices to estimate AGB and LAI across eight diverse quinoa varieties and multiple growth stages. We hypothesized that 3D mesh-derived metrics would be effective for quinoa because of its unique architectural diversity and cultivation practices that allow clear 3D mesh generation in individual plants. Genotype-specific patterns in estimation accuracy were observed, highlighting the impact of genetic diversity on the effectiveness of remote sensing. Temporal analysis revealed varying model performance across growth stages, with mid-season yielding the highest accuracy for AGB estimation. Cross-validation demonstrated model robustness, with plant volume-based models achieving R² = 0.883 (RMSE = 67.293 g/m², MAE = 39.262 g/m², RE = 21.015 %) for AGB and COM-based models reaching R² = 0.932 (RMSE = 0.053 m²/m², MAE = 0.044 m²/m², RE = 17.694 %) for LAI. Although underscoring the need for adaptive modeling approaches, our fusion models combining 3D mesh structural, multispectral, and RGB features achieved high estimation accuracy (R² = 0.919 for AGB and 0.946 for LAI) and remained robust against genotypic differences and temporal dynamics that typically affect individual features. These findings demonstrate the effectiveness of UAV-based remote sensing for nondestructive quinoa monitoring.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.