ESTIMATING YIELDS OF ROOT CROPS VIA OBJECT COUNTING USING DEEP LEARNING

Abstract

As crop predictions can assist in decision-making regarding crop selection, time period, duration, etc., there is a need for an automatic and accurate methodology for predicting yields in large agricultural areas. This study aimed to determine the pattern of distribution of yields using the videos of sugar beets captured from a camera attached to beet harvester. Deep learning networks for object detection, such as YOLOv2, YOLOv3, and EfficientDet, were employed to use the videos to automatically detect sugar beets on the conveyor belt of the harvester. The mean average precision of sugar beet detection with these detectors was over 0.95, indicating the high accuracy of object detection. Further, identical sugar beets in consecutive video frames were identified by object tracking using Kalman filter. This aided in the accurate counting of sugar beets flowing on the conveyor belt. The counting was performed with an F-value of over 0.95. For example, 840 of 847 sugar beets were properly counted. Also, the weight of each sugar beet detected by the deep learning network was estimated using regression technique, based on its long and short axes. The correlation between the size and weight of sugar beets was computed prior to the experiment on site, which enabled to predict the total weight of the harvested beets. The error of yield prediction using YOLOv3 and EfficientDet was about 3%. Based on the predicted yields of sugar beet that was harvested, the geospatial information was associated to each of the detected sugar beet as the camera recorded the GPS information while recording videos. Finally, the distribution pattern of sugar beet yield was visualized and the variations were analyzed.