Development of a System for Automated Control of Planting Density, Leaf Area Index and Crop Development Phases by UAV Photos

The paper describes the results of research into the possibility of creating a system for automated control of crop development indicators using UAV images and neural networks. Such architectures as YOLOv5, U-Net and FPN (Feature Pyramid Network) are considered. It is found out that YOLOv5m neural network is able to qualitatively detect plants on a sample of 363 images with an accuracy AP@.5:.95 of 92.4 %. The detection efficiency can be improved both by adding original images to the training sample and by increasing it artificially. Plant detection will help to determine the density of crops. It is shown that the calculation of the area of segmented pixels in an image can be used to determine crop development and leaf area index. FPNs are used for segmentation and it is obtained that FPNs are on average 1.05% more accurate than Unet and the average mIoU across the three classes is 1.26% larger.