Estimation of water stress in maize cultivation utilizing thermal and multispectral imaging from UAVs with machine learning algorithms in Lambayeque, Peru

• Autores: Cruz Grimaldo, Camila Leandra, Vilca Gamarra, Cesar Francisco, Millan Ramírez, José Edwin, Chumbimune Vivanco, Sheyla Yanet, Llanos Carrillo, Cristina, Vera Díaz, Elvis, Agurto Piñarreta, Alex Iván, Quille Mamani, Javier, León Dextre, Hairo Alexander
• Formato: artículo
• Fecha de Publicación: 2026
• Institución: Instituto Nacional de Innovación Agraria
• Repositorio: INIA-Institucional
• Lenguaje: inglés
• OAI Identifier: oai:repositorio.inia.gob.pe:20.500.12955/3018
• Enlace del recurso: http://hdl.handle.net/20.500.12955/3018; https://doi.org/10.4995/raet.2026.23671
• Nivel de acceso: acceso abierto
• Materia: Crop water stress index (CWSI); Machine learning; Precision agriculture; Thermal image; Vegetation Index; Índice de estrés hídrico de los cultivos (CWSI); Aprendizaje automático; Agricultura de precisión; Imagen térmica; Índice de vegetación; https://purl.org/pe-repo/ocde/ford#4.01.01; Zea mays; Maíz; Maize; Estrés Hídrico; Water stress; Agricultura de precisión; Precision agriculture; Teledetección; Remote sensing; Vehículo aéreo no tripulado; Aerial vehicles; Riego; Irrigation.

Maize (Zea mays L.) is a fundamental cereal in global food security, but its vulnerability to water stress compromises its productivity and threatens food availability. This study analyzed the relationship between the crop water stress index (CWSI), obtained from thermal images captured by the Zenmuse H20T camera, and various vegetation indices derived from the MicaSense RedEdge-MX Dual. The analysis included machine learning (ML) models such as random forest (RF), k-nearest neighbors (KNN), and gradient boosting regression (GBR). The results showed that RF was the most accurate model for predicting CWSI in maize, with a coefficient of determination (R²) of 0.80, a root mean square error (RMSE) of 0.13, and a mean absolute error (MAE) of 0.09. KNN achieved an R² of 0.78, an RMSE of 0.13, and an MAE of 0.09, while GBR reached an R² of 0.79, an RMSE of 0.14, and an MAE of 0.10. The red band (668 nm) played a crucial role in RF (70.69%) and GBR (50.92%), whereas in KNN, the simple ratio (SR) index showed the highest importance (36.40%). These findings confirm the superiority of ML models over traditional regression approaches for estimating CWSI in maize. Despite the satisfactory results, the algorithms underestimated CWSI values derived from thermal images, which highlights the need to refine these models to improve their accuracy in future agricultural applications.