Wheat head identification through a smartphone camera using YOLOv5 object detection algorithm

In this study, we have applied an advanced technique in the field of computer vision, the YOLO (You Only Look Once) algorithm to accurately detect, count, and analyze wheat spikes from the images captured through a mobile phone at real field conditions. Wheat head detection can provide the main source of information such as density, presence of awns, health, maturity stage, and size to agronomists and farmers. It helps to make optimal grain management decisions such as yield estimation and prediction. An experiment was performed to examine the utility of the currently available open-source toolkit in computer vision to build and employ Convolutional Neural Network (CNN), applied to solve various agricultural and food production challenges. Using the open-source code available on the official Ultralytics GitHub page https://github.com/ultralytics/yolov5 , the typical process of computer vision was implemented with the training dataset of 360 images taken from Global Wheat Head Dataset (GWHD) and validation dataset of 700 mobile phone images captured from the educational farm of the University of Padova, Italy. Developed and trained Convolutional Neural Network with manually annotated image dataset to employ deep learning-based image recognition technique for detecting wheat spikes under varying field conditions. Explained technical details of employing object detection algorithm YOLOv5 for agronomy applications. Discussed the overall precision and F1-Score of the developed model to predict actual wheat spikes present in the images. A brief introduction to different evaluation matrices to improve the accuracy and performance of object detectors. Analyzes the comparisons of the developed model with manually recorded spike counting from the farm to estimate the best timing and angles to capture pictures in mobile phones. The result indicated that images captured around 6 pm with a 15-20° angle show a better understanding of finding a relationship between data features and target labels, making decisions, and evaluating their confidence from the training wheat head dataset provided. It can be concluded that a well-trained model can detect spikes from the given images, the accuracy of the detection is highly dependent on the size and quality of the dataset used for training the model. Furthermore, the future scope of implementing AI and object detection approaches were discussed to solve the problem of precise identification of the plant phenotypic problems.