Estimation Of Wheat Tillers And Ear Numbers Based On Terrestrial Laser Scanning

As important phenotypic parameters of yield,real-time assessment of the number of tillers and ears is of considerable contribution to monitoring the growth of wheat populations or as a crucial phenotyping indicator for screening of highly productive cultivars in crop breeding.The terrestrial laser scanning can be used to obtain detailed and accurate three-dimensional structure information of a crop regardless of the ambient light conditions—an advantage over the use of optical remote sensing.In this study,field trials with different varieties,planting densities,and nitrogen levels were carried out in the same experimental area for two consecutive years.The manual measurements were performed at six key growth stages of wheat(stages of tillering,jointing,heading,flowering stage,early filling and end filling),to simultaneously collect the terrestrial Li DAR data and the tiller and ear numbers of the crop.Based on the previous researches,an improved tiller-counting algorithm was developed here.And two ear number estimation algorithms were constructed considering the morphological characteristics of the stem and leaf of the crop.This was a useful attempt to reduce the influence of noise and occlusion on the algorithm,which would promote the application of the terrestrial laser scanning in agriculture.Firstly,we improved the tiller-counting algorithm published by previous researchers.The previous algorithm combined the adaptive layering algorithm with the clustering algorithm in estimating the number of tillers.However,the accuracy of the algorithm was not satisfactory due to the negative influence of the occlusion and noise.In this study,we proposed a mean shift algorithm based on voxel interpolation(MSBVI)aimed at these two problems,which was mainly divided into three steps.(1)The Kalman filter algorithm was introduced in the preprocessing step to remove excess noise in order to improve the quality of the initial data.(2)After the voxelating step,we established a formula between the gap fraction and the point cloud density of voxels.Then we calculated the gap fraction of voxels in the neighborhood of the unsampled voxel,and substituted it into the relationship to obtain the point cloud density of the unsampled voxel.The interpolation of the missing voxels in the canopy was completed so as to reduce the negative influence of occlusion on the algorithm.(3)The mean shift algorithm was applied to clustering the canopy after interpolation to obtain the number of ears.The results showed that the improved algorithm slightly raised the R~2,from 0.65 to 0.68 before interpolation,and further to 0.69 after interpolation.An obvious improvement could be observed,at the same time,in RMSE,from 102 tilllers/m~2 to 88 tilllers/m~2 before interpolation,and further to 79 tilllers/m~2 after interpolation.In addition,for the experimental conditions(the tillering stage,the 25cm planting density and the scattered plant type)with lower estimation accuracy in the previous algorithm,the interpolation algorithm presented more significant effect in improvement.Secondly,we estimated the number of ears in the field based on the terrestrial laser scanning.By referring to the related application experience in forestry,we developed two ear-counting algorithms.One was a density-based spatial clustering of applications with noise algorithm based on the normal density-based spatial clustering(NDBSC),and the other was a voxel-based regional growth algorithm(VBRG).NDBSC was mainly divided into two steps:(1)Based on the different curvatures of the wheat leaf and stem,NDBSC was used as a threshold to segment the point cloud of the leaf and the stem(ear).(2)After the segmentation,the separated stem(ear)part was clustered using a density clustering algorithm,and the number of ears was calculated.VBRG was mainly divided into three steps:(1)The discrete point cloud was voxelated.(2)The kd-tree was used to construct a topological relationship between voxels to make them into connected regions;(3)The region growing algorithm was utilized to detect ears.The results showed that the overall performance of NDBSC(R~2=0.70,RMSE=72 ears/m~2)was better than that of VBRG(R~2=0.64,RMSE=114 ears/m~2).Compared with the latter method,the former one,in which points were regarded as the basic units of operation,obtained more detailed information.Therefore,NDBSC was more suitable to be used on the field wheat.To sum up,the general performance of the terrestrial laser scanning applied in this study to the estimation of the numbers of tillers and ears proved that it had the ability to obtain effective yield phenotypic parameters of crops.In the future,it will play a greater role in agricultural applications,ushering agricultural phenotype into a three-dimensional era.