Estimating and Mapping Rice Yield Using UAV-Hyperspectral Imager based Relative Spectral Variates

Crop yield is important for national and regional food production, food trade and food security. Traditional yield estimation by satellite remote sensing is limited by many factors such as spatiotemporal resolution and number of bands. UAV imaging hyperspectral technology has been widely applied to modern intelligent agriculture and precision agriculture with its advantages of high spatial and temporal resolution, rich band number and the combination of image and spectrum It is possible to estimate crop yield accurately. The multi-temporal vegetation indices for yield estimation are obtained with different illumination conditions, atmospheric conditions and background values, the differences in these external conditions may result in errors in vegetation indices. Therefore, using these multi-temporal vegetation indices which containing these external conditions for yield estimation is likely to cause errors. To address this problem, this study proposes the concept of “relative spectral variables” and “relative yield” to estimate rice yield using multi- temporal relative variables. Firstly, the bands obtained from hyperspectral imager are combined to establish the Relative Normalized Difference Spectral Index(RNDSI) and the optimal RNDSI are selected for different growth stages. Then, the optimal models of rice yield estimation with different growth stage combinations are determined and validated. The results shows that multiple linear regression model consisting of tillering stage RNDSI_{784, 635]}, jointing stage RNDSI_{807, 744]}, booting stage RNDSI_{784, 712]} and heading stage RNDSI_{816, 736]} is the optimal models for rice yield estimation with R² of 0.74 and RMSE of 248.97 kg/ha. This model is validated and the result is acceptable with average relative error of 4.31%. In conclusions, the relative vegetation index and relative yield can be applied to the pixel-level yield estimation by remote sensing. Besides, the rice yield distribution map is drawn based on the model, which represents the differences of rice yield at different filed positions. The map may be used to carry out precise field management.