Comparative Analysis for Remote Sensing of Canopy Nitrogen Content in Rice Based On Hyperspectral Measurements.

Assessment of canopy nitrogen content (CNC) is an important basis for growth diagnosis, precision management, and yield prediction in rice crop. High-resolution spectral reflectance measurement (Hyperspectra) has been suggested to have significant roles in assessment of crop variables, especially biochemical components (e.g., chlorophyll content) and physiological functioning (e.g., light use efficiency) as well as biotic and abiotic stresses (e.g., water deficit, disease infection). Nevertheless, optimal use of hyperspectra for each target variable has not been well examined. Here, we explored accurate and robust spectral indices or multivariable methods for assessment of CNC in rice based on multi-site measurements by ground-based and airborne sensors. Hyperspectral reflectance data were obtained using ground-based (ASD-FSFR; Li-Cor-L1800) and airborne (CASI-3) sensors. The range of nitrogen concentration, canopy nitrogen content, and biomass was 0.99-3.58 %, 0.31`16.52 g m<sup>-2</sup> and 8.9`709.8 DWg m^-2, respectively. Predictive methods compared were 1) simple spectral indices such as Normalized Difference Spectral Index NDSI(i nm, j nm) and ratio vegetation indices RVI(i nm, j nm) using reflectance spectra and first derivative spectra (FDi nm), 2) partial least squares regression (PLS) using the whole hyperspectra, and 3) PLS with waveband selection (IPLS), as well as 4) a range of previously proposed vegetation indices. When spectral reflectance at some limited number of discrete wavebands can be measured, NDSIs using two wavebands at the 700-850 nm region proved to be best among the thorough combination of two wavebands. If FD spectra can be derived, the ratio index using FD at red-edge and green wavebands would be most promising in predicting CNC, the predictive ability of which was better or comparable to that of PLS or IPLS approaches. IPLS using smaller number of wavebands showed higher predictive ability than PLS using the whole wavebands.