Comparison of different imaging sensors of Unmanned Aerial Vehicle (UAV) for wheat yield prediction
Wheat is the third-largest field crop in the U.S., many corresponding strategies can be made in advance if wheat yield can be predicted before the harvest time. Three different imaging sensors (RGB, RG-NIR and hyperspectral imaging) mounted on the Unmanned Aerial Vehicles (UAV) were used to predict wheat yield in this study. RGB camera could provide three bands (R, G and B channels) and RG-NIR could provide another three bands (R, G and NIR channels), while hyperspectral imaging could produce hundreds of wavebands in visible and near-infrared ranges (395-885 nm). A total of 252 (14×18) sub-plots in the field (44.991887N, 93.186792W) were imaged by RGB and RG-NIR sensors on 6/21/2016, while 220 (22×10) sub-plots in another field (44.989123N, 93.185543W) were imaged by hyperspectral imaging sensor in June and July 2017. RGB and RG-NIR images were taken using DJI Phantom Professional 3 drone, and hyperspectral images were taken by Pika II hyperspectral imaging camera mounted on DJI Matrice 600 Pro drone. After calibration, the background (i.e., soil and weed) was deleted, resulting in only the sub-plots were studied. Each sub-plot was treated as one region of interest (ROI), and the features (i.e., R, G, B, R-G-B, R-G, R-B, G-B, G/R, B/R, B/G, (R-G-B)/(R+G+B), (R-G)/(R+G+B), (R-B)/(R+G+B), (G-B)/(R+G+B), R/(R+G+B), G/(R+G+B), B/(R+G+B), mean, variance and entropy) extracted from RGB images and the spectral reflectance extracted from hyperspectral images were treated as X variables for modeling (linear and non-linear). The NDVI image were calculated using both RG-NIR and hyperspectral images. Then, the significant spectral wavebands were also identified to establish the prediction models. Hyperspectral imaging performed better than RG-NIR and RGB, and RG-NIR could produce better results than RG-NIR. For different date of hyperspectral images, each prediction model obtained the coefficient of determination (R2) higher than 0.6, and the best prediction date was also identified. The overall results demonstrated that imaging information collected by UAV could predict the wheat yield.
