Remote sensing of rice yield using UAV-derived SAVI and supervised machine learning models in Tropical Lowlands

Ysuiza-Perez A.; Perez-Tello M.; Goicochea-Pinchi D.; Vega-Herrera S.; Rios-Rios R.; Dominguez-Yap P.; Garcia L.; Barrera-Torres C.; Oliva-Cruz C.; Santillán-Gonzáles M.; Arratea-Pillco D.; Alejos-Patiño I.W. · Scientia Agropecuaria (2026) · DOI: 10.17268/sci.agropecu.2026.034

Resumen

Accurate estimation of rice productivity at the sub-field level is still a major challenge in tropical agroecosystems, mainly because of the high spatial variability and the limits of traditional monitoring methods. This study looked at how well the soil-adjusted vegetation index (SAVI), pulled from multispectral images taken by UAVs, could separate productive and non-productive zones in rice fields under tropical lowland conditions in San Martin, Peru. We used a randomized complete block design across two locations with three rice varieties and captured multispectral images at key phenological stages using UAV platforms. Field yield came from georeferenced destructive sampling—we adjusted grain weight to standard moisture and expressed everything in t ha^-1. Based on those actual measurements, we set threshold criteria to classify zones as either productive or non-productive. SAVI values were then extracted and fed into supervised classification models: logistic regression, support vector machine (SVM), k-nearest neighbors (KNN), random forest, and decision tree. The results showed that SAVI values between 0.50 and 0.70, typically lined up with productive zones, while 0.30 to 0.50 corresponded to non-productive areas. Logistic regression and SVM came out on top with overall accuracy around 88.9%, F1-scores above 92%, and pretty balanced sensitivity and specificity. These findings suggest that combining SAVI with supervised machine learning offers a solid, practical way to map rice productivity spatially. The approach looks promising for supporting intra-field monitoring and helping make better agronomic decisions in tropical rice systems. © Derechos de autor 2026 Scientia Agropecuaria. Esta obra está bajo una licencia internacional Creative Commons Atribución-NoComercial 4.0.

Instituciones: Universidad Nacional Autónoma de Alto Amazonas; Universidad Nacional Hermilio Valdizán