Abstract Construction And System Development Of Cotton Seedling Monitoring Model Based On UAV Remote Sensing

[Objective] To achieve timely seedling inspection and replenishment as well as monitoring of cotton seedling growth conditions during the seedling stage,and further provide a guarantee for the late production management and yield formation of cotton.(1)To construct a detection and counting model based on UAV multispectral images throughout the seedling stage.(2)Construct a growth index monitoring model based on UAV RGB and multispectral images.(3)Conduct comprehensive evaluation based on seedling parameters and growth indicators,and finally complete the design and implementation of the seedling growth monitoring system.[Methods](1)Firstly,multispectral images of cotton seedlings in red,green and NIR bands were acquired six times(T1 to T6 denote 10,16,20,22,24 and 27 days after sowing,respectively).Band synthesis and cropping were performed followed by model training based on Python 3.8 using three deep learning methods,YOLOv5,YOLOv7 and Center Net,and finally an optimal model that can be applied to seedling detection and counting throughout the seedling period was determined from the six acquired datasets.(2)Secondly,RGB and multispectral images of cotton seedlings were acquired,and samples were collected to determine growth indices for three times.Based on the extracted vegetation indices and texture features,feature screening was performed using correlation method(Corr),maximum information coefficient method(MIC)and random forest method(RF),and monitoring models were constructed using multiple stepwise regression(MSR),KNN regression(KNN)and extreme random forest regression(ET).Finally,the comprehensive evaluation was completed based on seedling growth parameters using the Entropy-weight-TOPSIS method.(3)Finally,the seedling growth monitoring system is designed based on technologies such as Python,My SQL and ECharts visualization tool.[Results](1)In the six collected datasets,the detection and counting results of T2-T5 were better,indicating that the images collected at 16-24 days after sowing for cotton seedling detection and counting were more effective.When testing the data collected at different times,YOLOv7 showed better detection and counting results,and the comprehensive performance of cotton seedling detection and counting was better in the T4 dataset.T5 showed the best results in the T4 dataset with Precision,Recall and F1-Score of 96.1%,95.9%and 96.0%,respectively,and R~2,RMSE and R~2,RMSE and RRMSE were 0.94,3.83 and 2.72%,respectively.(2)The growth index monitoring model constructed based on the combination of vegetation index and texture features extracted from RGB and multispectral images was better than the monitoring model constructed from vegetation index or texture features.Among them,the optimal estimation model for plant height was Corr＿ET with the training set R~2,RMSE and RRMSE of 0.87,1.2081 and 15.09%,respectively,and the validation set R~2,RMSE and RRMSE of 0.88,1.1971 and 14.84%,respectively.The optimal estimation model for above-ground biomass was RF＿ET with training set R~2,RMSE and RRMSE of 0.82,0.0170 and24.60%,respectively,and validation set R~2,RMSE and RRMSE of 0.73,0.0183 and 26.38%,respectively.The optimal estimation model for water content was RF＿ET with the training set R~2,RMSE and RRMSE of0.84,0.0119 and 1.40%,and the validation set R~2,RMSE and RRMSE of 0.81,0.0137 and 1.61%,respectively.(3)A seedling growth monitoring system was developed to achieve rapid estimation of growth indicators using existing models.[Conclusion] In this study,the seedling detection and counting of cotton at the seedling stage was feasible using multispectral images and deep learning technology.16-24 days after sowing,the detection and counting of cotton seedlings were better,and based on the YOLOv7 target detection algorithm,the T4 dataset achieved fast and accurate seedling detection and counting throughout the cotton seedling stage.The monitoring model constructed using RGB and multispectral images can better estimate plant height,above-ground biomass and water content.The developed seedling growth monitoring system can quickly perform growth index estimation.