High-throughput Phenotype Monitoring Of Rice Field Canopy Structure Using UAV RGB Images And Deep Learning

Rice is one of the most important food crops.Accelerating the cultivation of rice cultivars adapted to the future climate is the key to coping with potential food crises.High-throughput monitoring of canopy structure dynamics under field conditions is of great significance for improving the efficiency of new product selection.Among them,Green Fraction（GF）and Green Area Index（GAI）are the most important canopy structure characteristics that characterize the light interception capacity and growth rate of the population.This research aims to explore high-throughput and high-precision methods for estimating rice GF and GAI,taking into account the effects of sowing date,cultivars,fertilization and climate factors on the growth of rice,and to carry out rice field experiments.Using UAVs supplemented by handheld devices,mounted with high-definition RGB cameras,take images of key growth periods of rice from a 45°zenith angle.The exploration of related issues is carried out on two aspects:accurate calculation of GF of rice field images and high-throughput and high-precision estimation of rice GAI.Firstly,in the accurate calculate GF of the rice field image,the data source is the RGB image of rice obtained by the UAV from 4 observation heights（25 m,35 m,50 m,70 m）supplemented by the handheld device from the near ground（2 m）.GF is the segmentation accuracy evaluation index,and the following two researches have been carried out:1)Study the impact of the training set composition on the segmentation accuracy and robustness of the U-net model;2)Evaluate the combination of the U-net model and Ex G（Excess Green）combined with the Otsu automatic threshold method,and calculate the accuracy of the GF of images obtained at different heights.The results show that:compared to using a single spatial resolution image as the training set,the output model obtained by training U-net with a training set containing UAV low-altitude（25 m）images and near-ground images has the best robustness（RMSE=0.057）.Using the U-net model to segment near-ground images（GSD=0.07 mm）and UAV low-altitude images（GSD=0.17 mm）both have higher accuracy,and the RMSE is 0.043 and 0.061,respectively.As the spatial resolution of the image decreases,the segmentation accuracy of the U-net model decreases.The calculation error of Ex G algorithm is larger than U-net on different test sets,and with the increase of GF,the error of Ex G algorithm will increase greatly,but the segmentation accuracy of U-net model is hardly affected by the change of GF.In addition,this study proves that the accuracy of using U-net model to segment low-altitude UAV images can reach the accuracy of segmenting near-ground images,thereby pave the way for replacing handheld devices with UAVs for high-throughput and high-precision monitoring of rice canopy GF.Secondly,in the high-throughput and high-precision estimation of rice GAI,the data source is the RGB image of rice obtained by UAVs from 4 different heights（25 m,35 m,50m,70 m）and the measured rice GAI in the field.The Ex G median value of rice canopy(Ex G_med)in UAVs images is not sensitive to changes in spatial resolution（or flight height）and has a strong correlation with GF.On the basis of which Ex G_med are correlated with GF calculated from different flight heights,we constructed a regression model（or called High-Low-Altitude Correction Model,HLACM）,and then calculate the GF of each rice cultivars in the high-altitude image,and finally construct three GAI estimation models:an exponential regression model of 1-Ex G_med and GAI;logarithmic regression model and Poisson model based on the gap fraction（1-GF）.The results show that with the help of HLACM with different altitudes,it is possible to avoid directly calculating the GF of the rice field from the high-altitude low-resolution image,thereby improving the accuracy of the GF calculation.The accuracy of the model that uses the gap fraction to estimate GAI is relatively high,but it is susceptible to clumping effects.Among them,the Poisson model is based on the radiative transfer theory and has strong robustness.It can obtain dynamic GAI information;the logarithmic model can reduce the overall deviation of GAI estimation of multiple growth periods of rice.In general,this paper systematically analyzes the current shortcomings in obtaining field rice GF and GAI.With the help of the deep learning U-net model,we extracted rice field GF with high precision;and by constructing a correction model,we further realized the high-throughput and high-precision estimation of field rice GAI using UAVs.This research idea can be extended to crops whose canopy structure is similar to rice.This research will help promote the development of accurate and efficient estimation of crop GF and GAI,provide scientific basis for large-scale crop phenotypic parameter monitoring and research,and it is of great significance for optimizing cultivation management and breeding selection.