In this study, a spatial analysis of agronomic and remote sensing data is carried out to derive accurate rice crop yield estimation. The variability of a series of vegetation indices (VIs) was calculated from remote sensing data obtained via a commercial UAS platform (e-Bee) at four dates (per stage of development), and the development of estimation models was conducted. The study area is located in the region of Chalastra (municipality of Thessaloniki, North Greece) and the primary data were obtained during the 2016 growing season. These data include ultra-high resolution remote sensing multispectral images of 18 plots totaling 58 hectares of Ronaldo and Gladio rice crop varieties, 97 sample point data related to yield, and many other pieces of information recorded in the producer’s field log. Ten simple and compound VIs were calculated, and the evolution of their values during the growing season as well as their comparative correlation were studied. A study of the usability of each VI was conducted for the different phenological stages of the cultivation and the variance of VIs and yield; the more correlated VIs were identified. Furthermore, three types of multitemporal VI were calculated from combinations of VIs from different dates, and their contribution to improving yield prediction was studied. As Ronaldo is a Japonica type of rice variety and Gladio is Indica type, they behave differently in terms of maturation time (Gladio is approximately 20 days earlier) and the value of every VI is affected by changes in plant physiology and phenology. These differences between the two varieties are reflected in the multitemporal study of the single-date VIs but also in the study of the values of the multitemporal VIs. In conclusion, Ronaldo’s yield is strongly dependent on multitemporal NDVI (VI6th July + VI30 Aug, R2 = 0.76), while Gladio’s yield is strongly dependent on single-date NDVI (6 July, R2 = 0.88). The compound VIs RERDVI and MCARI1 have the highest yield prediction (R2 = 0.77) for Ronaldo (VI6th July + VI30 Aug) and Gladio (R2 = 0.95) when calculated in the booting stage, respectively. For the Ronaldo variety, the examination of the multitemporal VIs increases yield prediction accuracy, while in the case of the Gladio variety the opposite is observed. The capabilities of multitemporal VIs in yield estimation by combining UAVs with more flights during the different growth stages can improve management and the cultivation practices.
Spatial Analysis of Agronomic Data and UAV Imagery for Rice Yield Estimation
