Regression-Based Models to Predict Rice Leaf Area Index Using Biennial Fixed Point Continuous Observations of Near Infrared Digital Images

Abstract

A weatherproof digital imaging system for the near infrared band (NIR, 820–900 nm) was positioned 12 m above a 600-m2 rice field. During the 2008 and 2009 paddy rice seasons, the system automatically logged images at 10-min intervals throughout the day. Radiometric corrections for the NIR images utilized a solar irradiance sensor and prior calibrations to calculate 0900–1500 JST daily-averaged reflectance factors (DARF). Prior to heading, empirically derived equations for predicting leaf area index (LAI) using the 2008 DARF values in NIR, the cosines of angles between the view and the planting row directions, and between the view and the meridian directions were verified with the 2009 data set. Transformation of a model variable by arcsine square root function improved the performance of the LAI prediction by reducing the errors and bias at low LAI values. Adding variables to incorporate lateral angular components to the horizontal viewing angular parameters hardly affected the overall performance of the models and did not reduce variation. This was probably because the height and position of the camera system were the same in successive years. In-plot means of two or four predicted values in each plot reduced the root-mean square error 30%. These results indicate that radiometric NIR images derived using a fixed-point observation system can accurately predict LAI and the simple multiple linear regression equations developed for a given year can be used the following year without in-situ recalibration.