Rice yield estimation at pixel scale using relative vegetation indices from unmanned aerial systems

Abstract

Timely and accurate prediction of rice yield information is closely related to the people’s livelihood, which has been attached great importance by all levels of government. Satellite remote sensing provides the possibility for large-scale crop yield estimation, but they are usually limited by spatial and spectral resolution. Unmanned Aerial Vehicles (UAV) remote sensing with hyperspectral sensors can obtain high spatial-temporal resolution and hyperspectral images on demand. Generally, time-series Vegetation Indices (VIs) are used for estimating grain yield. But multi-day vegetation indices may be affected by different background and illumination condition, so the differences between vegetation indices may include the effects induced from external condition, which will pose a negative effect on the accuracy of crop yield estimation. Therefore, in this study, the relative vegetation index and relative yield were proposed and used to estimate rice yield at pixel scale. And the optimal growth stages for crop yield estimation would also be determined. Hyperspectral images of critical rice growth stages at tillering stage, jointing stage, booting stage, heading stage, filling stage, ripening stage were obtained from July 28 to November 24 in 2017. Firstly, all possible two-band combinations of discrete channels from 500nm to 900nm was used to create Relative Normalized Difference Vegetation Index (RNDVI). Then the best RNDVI at different growth stages were determined for rice yield estimation. Finally, different combinations of growth stages were tested to obtain the optimal combinations for yield estimation. These models were validated at pixel scale using the measured yields. The result shows that four-growth-stage model with RNDVI[635, 784] at tillering stage, RNDVI[744,807] at jointing stage, RNDVI[712,784] at booting stage, RNDVI[736,816] at heading stage with the multiple linear regression function gain a higher R2 (0.74) and lower RMSE (248.97kg/ha). The mean absolute percentage error of estimated rice yield of 4.31%. Results shows that the yield estimations at pixel scale with relative vegetation indices were acceptable. In the study, a yield estimation method with relative vegetation indices is proposed and the optimal growth stage combinations for rice yield estimation were determined. This study explores the possibility of yield estimation at pixel scale using hyperspectral images from UAV platform, which will further improve the method system for remote sensing of yield estimation.