Potential of Hyperspectral Imaging for Constructing a Year-invariant Model to Estimate the Nitrogen Content of Rice Plants at the Panicle Initiation Stage

In this study, ground-based hyperspectral remote sensing was used for estimating the nitrogen content of rice plants (Kinu-hikari) at the panicle initiation stage. The resulting hyperspectral images were separated into two parts: (1) the rice plant and (2) others (e.g., irrigation water, soil background) using the equation of “GreenNDVI-NDVI”. RRICE was calculated as the ratio of the reflectance for the rice plant to that for the reference board. A partial least squares regression (PLSR) model was constructed based on the relationship between the reflectance of the rice plant and its nitrogen content. A PLSR with another parameter set (cumulative temperature) was also constructed in order to account for differences in weather conditions among years. PLSR models were evaluated for each year using leave-one-out cross-validation, RMSE ranged from ranged from 0.48 to 0.65 g/m2, and RE ranged from 7.0 to 15%. When each year's model was used to calculate the other years' data to determine the estimation power of a year-invariant method, RMSE values of mutual estimation were higher due to over- or underestimation (RMSE ranged from 0.49 to 3.95 g/m2, RE ranged from 8.0 to 85%). When a PLSR based on two years' data was fitted, RMSE ranged from 1.29 to 3.32 g/m2 and RE ranged from 21 to 43%. By contrast, in PLSR models incorporated both reflectance and cumulative temperature, RMSE and RE fell to less than 0.95 g/m2 and 13%, respectively. Thus, a combination of reflectance and temperature data was useful for constructing a year-invariant model to estimate rice nitrogen content.