Prediction of biomass and nutritional quality of tropical pastures using multispectral analysis and machine learning models

• Autores: Tafur Culqui, Josué, Atalaya Marin, Nilton, Gómez Fernandez, Darwin, Taboada Mitma, Víctor Hugo, Cruz Luis, Juancarlos Alejandro, Neyra, Henri, Anchayhua Torres, Janella Jelin, Quichua Baldeon, Rosalía, Sánchez Fuentes, Teiser, Olano Camán, Yadhira Milagros, Barrazueta Campos, Mauro Adel, Tineo Flores, Daniel, Goñas Goñas, Malluri
• Formato: artículo
• Fecha de Publicación: 2026
• Institución: Instituto Nacional de Innovación Agraria
• Repositorio: INIA-Institucional
• Lenguaje: inglés
• OAI Identifier: oai:repositorio.inia.gob.pe:20.500.12955/3153
• Enlace del recurso: http://hdl.handle.net/20.500.12955/3153; https://doi.org/10.1016/j.atech.2026.102229
• Nivel de acceso: acceso abierto
• Materia: Crude protein prediction; Predicción de proteína cruda; Dry matter; Materia seca; Extra trees; Random forest; Bosque aleatorio; Remote sensing; Teledetección; UAV multispectral imagery; Imágenes multiespectrales UAV; https://purl.org/pe-repo/ocde/ford#4.01.00; Biomasa, Biomass; Valor nutritivo, Nutritive value; Aprendizaje automático, Machine learning; Índice de vegetación, Vegetation index; Vehículo aéreos no tripulado, Unmanned aerial vehicles; Forraje, Forage

Determining pasture productivity and nutritional value through non-destructive approaches aimed at optimizing forage resource management and improving efficiency in livestock systems has become an urgent priority. In this context, the objective of this study was to evaluate the performance of machine learning models in predicting biomass production and the nutritional contribution of different pasture species, as well as to assess the role of vegetation indices (VIs) in these predictions. To this end, a multispectral sensor mounted on a DJI Matrice 350 RTK platform was used, together with agronomic, yield, and nutritional variables. The curated dataset was subsequently analyzed using linear and polynomial models, as well as tree-based algorithms and support vector machines. Model validation was performed using a group-constrained random partitioning scheme (Group Shuffle Split), with species considered as the grouping variable. Model interpretability was addressed through the SHAP (SHapley Additive Explanations) framework. The results indicated better predictive performance for yield-related variables compared to nutritional attributes. In particular, the Extra Trees model achieved the highest coefficients of determination (R²). SHAP analysis revealed that the Visible Atmospherically Resistant Index (VARI) contributed more strongly to yield-related predictions, whereas the Normalized Difference Red Edge (NDRE) showed a more consistent contribution to nutritional variables. In conclusion, these findings highlight the potential of integrating vegetation indices and machine learning models as effective tools for forage management, supporting informed decision-making in livestock production systems.