Flooded rice variables from high-resolution multispectral images and machine learning algorithms

Remote spectral detection via orbital, aerial or terrestrial platforms is considered a valuable tool for non-destructive real-time estimation of the Leaf Area Index (LAI), the status of plant N, and grain yield. In this context, this study aims to build predictive models from very high-resolution multispectral images as input variables with Machine Learning (ML) algorithms to generate indirect estimates of LAI, Narea, and grain yield for flooded rice culture. Multispectral images were acquired through a Sequoia & REG; camera aboard the Phantom 4 & REG; Pro platform, during five phenological crop stages. In addition to the spectral bands, nine vegetation indices were taken as predictors of the response variables derived from the site survey. The Spearman's test demonstrated a more significant correlation at the end of the vegetative stage (V7) and the beginning of the reproductive stage (R1) to predict the studied variables. Furthermore, the Support Vector Machine (SVM) models showed high fit and good generalization capability in flooded rice cultivation, reinforcing the excellent combination capacity between remote sensing via Remotely Piloted Aircraft Systems (RPAS) and machine learning in precision agriculture applications.