Infield peanut pod counting using deep neural networks for yield estimation

Abstract. Peanuts are the seventh most valuable crop in the U.S., with a farm value of more than $1 billion. Peanut breeders have a major role in improving the quality of the seeds used by the farmers as a way of maintaining the high productivity. Though, a large barrier for the breeders is obtaining yield data because it is labor intensive and time-consuming, which limits the scale of field trials. Computer vision and deep learning have been employed successfully in detection, tracking and segmentation in complex outdoors scenes. In this study, we evaluated the feasibility of predicting yield using video-derived pod counts. A pushcart imaging system was developed to collect side-view and top-view videos of peanut plants after digging. We processed each camera video independently through a pipeline using deep learning and classical methods to perform detection, counting and yield prediction. We used a pre-trained Mask R-CNN that is fine tuned to our peanut dataset. The correlation of video-derived peanut pod count was evaluated to predict yield data for research of plots of different genotypes. The results showed that the proposed pipeline was capable of accurately counting pods from different genotypes in complex dynamic scenes where multiple peanut plants are clustered together in an outdoor environment showing its capability of modeling the yield with a R² of 0.97 for the best view. The proposed high-throughput phenotyping system has potential to allow peanut breeders to test more genotypes in yield trials with improved efficiency.