Evaluation of Machine Learning to Estimate LAI Including Solar Radiation Condition at UAV Monitoring in Paddy Fields

Abstract

Leaf Area Index (LAI) is one of the most important indices for monitoring crop growth, but obtaining field measurements is laborious and time-consuming. In recent years as unmanned aerial vehicles (UAVs) have become a popular tool, it has been expected that estimation methods of LAI will be developed for UAV multispectral monitoring. However, our previous study suggested that solar radiation conditions distinctly affect observed values of canopy reflectance and vegetation indices. In this study, we examine several LAI estimation methods based on field measurement data for UAV multispectral monitoring under various solar radiation conditions. Eight paddy fields located in Sendai, Miyagi Prefecture, were selected for the field measurement of LAI and UAV multispectral monitoring. Compared with the LAI estimation based on non-linear regression using the enhanced vegetation index 2 (EVI2) as an explanatory variable, that based on multiple regression using EVI2, background reflectance and solar radiation conditions as explanatory variables improved root mean square error (RMSE) from 0.532 to 0.496 at LAI under 3.5, while it made RMSE worse from 0.929 to 1.120 at LAI over 3.5. This result suggests that the improvement of estimation accuracy by including solar radiation conditions is not consistent, probably due to the uneven data for the conditions. The LAI estimation based on support vector regression using the training data simulated by a radiative transfer model improved RMSE to 0.458 at LAI over 3.5. This implies that even data for the conditions provided by the radiative transfer model simulation properly trains the support vector regression, and improves the accuracy and robustness for LAI estimation.