Abstract
Remote sensing has been applied to estimate the protein content of brown rice in the field, but some uncertainty remains due to various error sources. Here, we investigated the major error sources to improve the accuracy in operational applications. Hyperspectral reflectance data were acquired using an airborne sensor over 100 km2 rice paddy areas. Grain protein content as well as plant conditions such as leaf color, height, and density was obtained by ground-based measurements. Correlation of the protein content with growth conditions,and reflectance spectra were analyzed using NDSI (Normalized Difference Spectral Index). The leaf color had a close and consistent relation with the protein content. Leaf color was closely correlated with reflectance in green to red spectral regions whereas protein content was less correlated especially in the red region. This difference was attributed to the higher sensitivity of red region to growth stage. A schematic model revealed that the protein content scarcely correlated with the planting date but closely with the fertilizing condition. Consequently, large variability of the growth stage due to different planting date was a larger error source in remote sensing of protein content. Earlier observation also resulted in lower accuracy because the large effect of planting date remained during the earlier ripening stages. The use of NDSI (green, near-infrared) instead of NDVI=NDSI (red, near-infrared) would improve the accuracy since it was less sensitive to the difference in growth stage.
